Hyun-Song Shin introduced his method back in the early 1990s in a series of obscure yet deeply impactful academic papers on the topics of insider trading and the favourite-longshot bias in horse racing markets. When I set about writing this article, I had no idea what he’d done in the intervening years, assuming he’d simply faded further into obscurity, holed up in some academic backwater, and thinking it was probably odds on that he was, in fact, dead. Imagine my surprise when a quick Google revealed that I’d got this one very, very wrong. Ladies and gentlemen, allow me to introduce you to the current Governor of the Bank of Korea:

Let’s start getting into the details by talking about the what. When a bookmaker posts odds on an event, they are telling us something about how likely they think each outcome is to happen - in other words, their true probabilities. But they are only telling us something. We don’t get to see these true probabilities directly - only the odds, which are the result of taking the true probabilities and applying a margin that earns the bookmaker their keep, pays their taxes, and protects against adverse selection - i.e. the possibility that their true probabilities are incorrect.

The margin is called the overround or - in US parlance - the vigorish or simply the vig. Hence techniques like Shin’s method are called overround removal or devigging methods. By undoing the margination, we can recover the bookmaker’s true probabilities - albeit never exactly and always with some degree of uncertainty. Different methods can therefore be judged by how well they do this: how accurate the resulting probabilities are.

Now let’s talk about the why. Many profitable trading strategies can be developed by treating one venue, instrument or counterparty as the source of truth and finding sufficiently large divergences from that source of truth to trade against. For example - respectively - deriving true probabilities from Betfair and finding overlays on Polymarket; using the Win market in horse racing to price exotics like quinellas and trifectas; and treating Pinnacle prices as the word of God sent down to Sinai. Arguments for this include: an imbalance in the proportions of informed versus dumb money; regional biases such as lopsided flow on national teams or the popularity of certain sports; and the teams of traders and quants that may be involved in setting the odds (depending on the bookmaker’s business model).

In all such instances we are still faced with the same question: how do we translate the observed odds into probabilities? This is why overround removal is so important and why your choice of method can make a meaningful difference to your trading performance. A better method means more accurate predictions and more accurate predictions means greater success in the hunt for positive expected value bets.

Shin derived his method by asking how a bookmaker should set their odds for a horse race if there is some proportion of bettors, denoted as z, that are “insiders” who know for sure which horse is going to win. The rest of the population of bettors, 1 - z, are uninformed, and their aggregate betting volume is assumed to be placed on each horse in proportion to the horses’ winning probabilities.

I can already hear your objections:

• I don’t bet on horses.

• If insider trading was that prevalent, it would be economically ruinous for bookmakers.

• The favourite-longshot bias is nigh on guaranteed to break the assumption that uninformed punters bet in proportion to the horses’ winning probabilities.

All very sensible. And all irrelevant. As professional gamblers, pragmatism trumps sophistication and good enough beats mathematical elegance. Or, in the immortal words of Euan Sinclair:

I am a trader. I am not a mathematician, financial engineer, or philosopher. My success is measured in profits. The tools I use and develop need only be useful. They need not be consistent, provable, profound, or even true.

So just think of z as a nuisance parameter, not a meaningful estimate of the proportion of insiders.

Indeed, Shin’s method has been shown, in the academic literature, to produce highly competitive predictions compared to alternative methods, on a range of different sports. See, for example, E. Štrumbelj’s paper “On determining probability forecasts from betting odds” or the thoroughly excellent “Adjusting bookmaker’s odds to allow for overround" by Clarke et al.

But let’s qualify that. Štrumbelj’s paper did not include the “power method” of Clarke et al. which the latter group found to perform consistently the best across a range of datasets and performance metrics. Clark et al. included no exchange odds, only studied tennis, horse racing and greyhound racing, used average bookmaker odds for the tennis dataset rather than focusing on, say, Pinnacle, and used obscure Australian and New Zealand tote data for horse racing and greyhound racing respectively. Their paper was also published close to a decade ago and some exciting new methods have been proposed in the past few years.

It is entirely plausible that different overround removal methods perform better in different contexts. Some things I think matter:

• The size of the overround. For every selection, the overround defines an interval within which the true probability lies. The bigger the overround, the bigger the interval and the harder that true probability will be to find. The choice of overround removal method will be more important for bookmaker odds than prediction markets and, in the latter case, for less liquid markets than those awash with money.

• Number of runners. A two-selection line market is completely different to a 25 horse cavalry charge.

• Distance from evens. If you trade line markets and specifically main lines then you’re going to be betting close to evens. There’s not much that overround removal can do here: the bookmaker has specifically picked the line so that the probability of either side is as close to 50% as possible. Potentially heavily lopsided moneylines are something else entirely.

• Hard-to-price outcomes. The canonical example of this would be draws in soccer which are notoriously difficult to predict. It makes sense that bookmakers would factor this in to how they distribute their margin across the three selections.

• Sports with tribal fans. As touched on previously, the popularity of sports teams can affect flows and also introduce a regional bias to those flows. While certain athletes in individual sports such as tennis may capture the public’s imagination, it’s unlikely to achieve the same level of dogmatism. And this applies even more so to a niche sport like horse racing - outside the very biggest events and once-in-a-generation horses.

• Number of winners. Some more esoteric markets such as the place market in horse racing pay out multiple winners. More mainstream examples include betting on which teams will make the playoffs or - particularly evocative of the zeitgeist - what words or phrases Trump will mention in August 2026. To the best of my knowledge, there’s been no prior work comparing the performance of overround removal methods on such markets and, indeed, some methods do not even work with multiple winners.

So let’s see how Shin’s method stands up - in 2026 - against the additive method, the multiplicative method, and the power method. You can refer to Clark et al. for the technical details. Missed the link? It’s here. But we’ll go further and also include (code available here: https://github.com/gotoConversion/goto_conversion) and Karl Whelan’s Shin-inspired method intended to more realistically model insider trading (code snippet provided below).

We’ll do this comprehensively, on five different datasets:

• Soccer moneylines

• Soccer totals

• Tennis moneylines

• Horse racing win markets

• Horse racing place markets

This way we cover, respectively:

• Three outcome markets

• Line markets - i.e. two outcome markets close to evens

• Two outcome markets with potentially highly skewed odds

• Many runner markets

• Markets with multiple winners

For the soccer and tennis data, we’ll use both Bet365’s and Pinnacle’s odds so we can answer the question of whether the size of the overround should determine the choice of method. For the horse racing data we’ll use Betfair’s Starting Price (BSP). This way we’ll cover a soft bookmaker, a sharp bookmaker and an exchange (what we used to call a prediction market in the good old days).

On each dataset, we’ll score the probabilities produced by each method using the Ignorance Score, more commonly referred to as the log-loss, the case for which is argued magnificently by Edward Wheatcroft in “Evaluating probabilistic forecasts of football matches: the case against the ranked probability score”.

Some final notes on the data before we get to the results. Soccer and tennis data were sourced from Joseph Buchdahl’s websites football-data.co.uk and tennis-data.co.uk respectively. I focused on data from the past 5 seasons but I want to highlight something of critical importance:

Pinnacle prices disappeared from the data towards the end of 2025. It’s been well documented that they’ve been locking down access to their odds. I therefore restricted the soccer and tennis analysis to matches where both Bet365 and Pinnacle odds were available. As an aside, frictions like these are compelling to a professional gambler: if you’re able to source valuable yet hard-to-find data that others aren’t then that’s a genuine competitive advantage.

Betfair horse racing starting prices - for both win and place markets - were sourced from https://promo.betfair.com/betfairsp/prices. This is a genuinely useful resource but the data is surprisingly messy. Paid subscribers can get access to a clean version of the data on request, along with all of the code used to produce the results.

We’ll start with the horse racing win markets. One interesting thing about BSP is that it often produces underround books - that happens on about 36% of markets. I don’t mention this in order to highlight a potential dutching strategy as this would be very challenging to take advantage of (exercise left to the reader) but rather because it has significant implications for turning the odds into probabilities.

There are some immediately fascinating takeaways here:

• The multiplicative method - the “standard practice” of simply dividing the inverse odds by the booksum - shockingly performs the best on both overround and underround books. When there is an overround, it’s small - about 1.1% - and that gives little room for the more sophisticated methods to distinguish themselves.

• For example, it turns out that there is good evidence of a favourite-longshot bias in these odds (not shown) and yet Shin’s method is unable to shine. It essentially runs out of road in the overround before it is able to fully adjust for the bias.

• For underround books, there’s not much to choose between the different methods. The very thing the methods are designed to remove doesn’t exist.

• Note that Whelan’s paper only provides the algorithm for two-runner markets so it has not been included in this comparison.

That’s a real banger straight off the bat: in the low overround environment of prediction markets, you’re sacrificing nothing by using the simplistic - and computationally cheap - multiplicative method.

Now let’s take a look at the 1X2 market in soccer. Remember that this time we are simultaneously comparing both bookmakers and methods and that reveals something striking.

Read more

This was my “reward” for winning that year’s Royal Statistical Society’s Statistics in Sport Section’s prediction competition (God what a mouthful). We’d been asked to submit predictions for all possible match-ups between the top 100 men and women tennis players at Wimbledon, a setup that should be familiar to anyone that’s filled out a March Madness bracket before. Some combination of my prior twelve years spent building statistical models for tennis, and my rare betting industry insight, carried the day and my predictions proved more accurate than the rest of my competitors.

Channelling my inner academic, I traded off that talk multiple times over the intervening years. I followed it up with a presentation to a meetup of professional gamblers hosted at Betfair HQ, and most recently gave an updated version for 2026 back in March at the London School of Economics (LSE) as part of the launch of this year’s competition, for which competitors have to predict the medals table for the Commonwealth Games.

I’d advertised the LSE talk here on Substack and remarkably a few dedicated - and London-based - readers turned up but I appreciated that the limits of spacetime would rule out attendance for most of my readership. To address this, I had intended to do a subsequent recording of the presentation myself but before I could get my shit together, I was serendipitously asked to present the material on The Corridor of Uncertainty podcast. Please enjoy:

This week was also the week that Chelsea Football Club hosted the annual “Smart Bash” conference, a thoroughly embarrassing riff on the much more famous US-based “Bet Bash”. It attracts a truly bizarre mix of clueless punters, industry veterans, horse-racing personalities, shadowy syndicates (that’s me), brokerages, and people trying hard to sell invariably unimpressive products. I first attended last year and the most complimentary thing I can say about it is that I chose to attend this year as well.

I say chose to attend because by Wednesday, the first day of the two-day conference, I was still unable to walk more than about twenty paces and literally incapable of sitting down. So rather than having spent £350 ($470) on a ticket to a conference of dubious quality, it was looking like I really had metaphorically set fire to the money. I knew several other people attending including my boss at Amelco who - in addition to sharing his own back-injury-related war stories - gave me blow-by-blow updates as the workshops unfolded. An initial session on how to legally structure gambling syndicates sounded the most practically useful, especially with us in the UK scratching our heads about how to access that sweet sweet prediction market loss flow without falling afoul of the law or the tax authorities on either side of the Atlantic.

Nothing got back to me concerning the second workshop entitled “Using Parlay Odds to increase your profits & sustain your accounts” but I hope we’re all well aware by now that same game parlays (SGPs) are the topic du jour, especially with Kalshi’s request for quote (RFQ) system upending the order and allowing the everyman to compete for those eye-watering bookmaker margins. As this workshop no doubt explained, parlays (accumulators) can also be used to compound expected value (EV) rather than margin, and also to disguise sharp behaviour by combining both positive and negative EV bets.

The verdict on the third session, on modelling NFL? Interesting but little immediately actionable and it confirmed the preconception that it’s a sport that’s hard to model. I’m reminded of the Risk Takers Podcast episode with Maple where he talks about how being a profitable NFL originator doesn’t make you the end boss of betting:

If you say, let’s say you’re like a super good NFL sides bettor, I would say maybe like maybe that’s three games a week and maybe you think you can make like 5% return on those three games a week.

That is 2.7 units for an entire season.

Too few games in a season combined with an incredibly competitive market? Don’t waste your time.

The final workshop session was a sales pitch for betunfair.io, an entirely unremarkable looking real time odds comparison website. They’re not the first to frame Pinnacle as the source of truth and they won’t be the last. As I’ve written about before in these pages, pricing is often the least important part of the professional gambling process, and being able to use a product like this to identify when some incredibly niche bookmakers are out of line with Pinnacle is useless unless you can get an account with said bookmakers and actually get down. There’s also a big problem with the framing, that we see manifesting on prediction markets: with hundreds if not thousands of aspiring bettors copying the Pinnacle odds - that may have a max bet size of, say, $5,000 - and posting their own liquidity, we end up with millions of dollars available in a single clip supported by a bookmaker that’s only willing to risk a tiny fraction of that. Everyone’s playing with fire.

I woke up on Thursday morning fully resigned to the fact that I’d only get to experience Smart Bash 2026 vicariously. In addition to my Amelco boss, I knew a few of the betcode lads attending. In keeping with the obsessive privacy swirling around the conference - the taking of photos and videos is strictly prohibited - I will refer to them as Mr. Spencer, Paddy and The Dinosaur. The feeling of FOMO was strong - indeed, that was a key motivation in wanting to attend both this year and last year. For an individual professional gambler, unless you’ve really hit the big time, a £350 conference ticket is a big ask when you don’t have a company or venerable academic institution paying your way. On the flip side, all it takes is one monetisable insight and the cost of admission can be earned back in a single afternoon if not a single race or match. I couldn’t stop thinking about the nuggets of wisdom I might be missing out on.

I was on a work call, pacing around my flat while discussing preparations for the World Cup. I love to pace while on a call, and it drives people crazy. It makes them nervous, I wear out their carpets, etc. etc. Suddenly it hit me: I’m pacing around my flat. I’m walking fine. It feels good. Maybe it would not be entirely insane if I was to head over to Smart Bash and at least make an attempt to salvage my week? The timing was tantalising. I’d miss the start but I’d make it in plenty of time to catch lunch with Mr. Spencer and co. then indulge in the afternoon’s panel discussions.

About an hour of sweaty Tube travel later, I was adorned with an embarrassing lanyard and catching the tail end of a veritable snooze fest of a panel discussion on the black market. Not the most encouraging of starts. But before I knew it, it was time for a livelier debate on the state of horse racing in the UK with Stephen Derbyshire in fine form, first setting his sights on the UK Gambling Commission (UKGC) before verbally skewering Zack Polanski. There wasn’t a whole lot of love in the room for the UKGC and their much maligned affordability checks, aside from a gentleman sat in the front row, you know the type, who I can only assume was mid-aneurysm when he decided to pipe up and push back against Derbyshire’s choice words for the UKGC’s representative at last year’s conference, conspicuous by his absence this time around. Luckily Derbyshire was having nothing of it and the febrile atmosphere quickly returned.

One of Derbyshire’s fellow panellists, one of the on course bookmakers, provided an interesting anecdote. I’m still trying to decide whether it is deeply profound or utterly devoid of meaning. He stated that you never want to be best price on a horse because that means you will be worst price on another. Does it? Does it really? Could you perhaps just reduce your overround? I often think about how we may all call ourselves professional gamblers yet we speak completely different languages. These old school horse racing fellows are as indecipherable to a UK exchange bettor such as myself as they would be to a US-based bettor rinsing DraftKings’ and FanDuel’s player prop offerings. At a superficial level it’s just the language of odds - fractional, decimal and American respectively - but it runs deeper than that. Maker vs taker, who your counterparty is, are you betting fixed odds, line markets or betting pools, do you trade price action or fundamentals. Heck, do you bet sports at all or are you betting on whether Trump will say the word “dog” in May 2026?

The less said about lunch, the better. But then came one of the most useful sessions - a live recording of The Smart Betting Club Podcast. It’s not a podcast that’s part of my usual rotation but a brief scan of the recent episodes shows an interesting and impressive range of podcast guests. They did put out a “Smart Bash preview” episode which you can check out if you’d like an alternative if somewhat biased take on the conference as a whole - and if you have nothing better to do with your valuable time. It looks like, as I’m writing this, that the live recording has yet to be published but presumably it’ll appear shortly. It turned out that each of the panellists had one key insight that I filed away for further reflection.

Jeevan Jeyaratnam talked about rules arbitrage, giving an example of goalscorer markets where one bookmaker would void only a single leg of a parlay if the player did not end up playing while another would instead void the entire parlay. Yet one of these bookmakers had blindly copied the odds of the other, without consideration of the rules. A bet is a contract and you’d be wise to know the terms of what you’re signing up for. We’ve also seen plenty of examples recently where prediction market participants have been burned by the rules - often not their fault but rather ambiguous wording on the part of the prediction markets themselves. There’s even a term for it: rules cuck.

It’s common knowledge that derivative markets, especially player props, can be attractive er… propositions due to weaker pricing and their potential to fly under the radar due to their association with uninformed punter behaviour. Varun Sudhakar, CEO of BetDEX, talked about how some of these are now attracting significant volumes, citing cricket player runs lines and NBA player props as two places where bettors seeking deep markets might wish to focus their efforts.

Finally, Phil Smith talked about his approach to ante-post betting in football (soccer). He’s made his own appearance on The Smart Betting Club Podcast which you can listen to in lieu of the live Smart Bash recording. I’m still not entirely sure what the fuck ante-post betting is in football as it’s a term normally associated with betting on blockbuster horse races in advance of knowing the definitive list of runners but from what I’ve managed to piece together, he’s talking about season long bets, also known as outrights, such as betting on who will win the Champions League, or who will finish in the top 6 of the Premier League. I’ve long considered these outright markets to be tantalising sources of potential profit, ever since being involved in just such a strategy at one of my previous jobs. They’re exactly the kind of markets where we might expect to find inefficiencies - they tie up money for a long time (the whole season) so they’re not especially popular and less interest generally means poorer pricing on the part of the bookmakers. But Phil made another excellent point about them that I’ve never heard articulated before: given they take so long to resolve, the bookmaker doesn’t realise you have an edge until it’s too late.

A coffee break offered a much needed pick-me-up and a chance to catch up with some old faces. Despite my cynicism, the conference does provide excellent networking opportunities. There were representatives from Betfair, Matchbook, BetConnect, Stats Perform, and Smartodds, to name drop just a selection of the attendees. But networking can be hard work at the best of times and my leg was starting to complain about the physical exertions I was putting it through. I stuck around for one more session, entitled “Prediction Markets vs Exchanges: Same Thing?” (conclusion: no shit Sherlock) but despite promising a lot it turned out to be incredibly dry with virtually no novel insight offered. A disappointing end to my Smart Bash 2026 experience, I thought, but there was no way I could physically make it through the evening’s planned networking session.

As I was gearing up to leave, I bumped into Carl, head of the Premium Team at Betfair and all round beautiful human being. We ended up having a great conversation. I thought Betfair would be really struggling with the intense competition from prediction markets but it turns out that the opposite is true. Perhaps it’s arbitrageurs or other cross-exchange activity but volumes are up, as are API user sign-ups. AI as a democratising force in software development may also be playing a role here. We chatted about Betfair’s own foray in prediction markets, Betfair Predicts, as well as their attempts to get SGPs onto the exchange, and I ended up trading some of these for the Champions League final. This I thought. This is what Smart Bash is all about. We all pay more than feels appropriate, they put on a mixed bag of content, we drink shit coffee and eat shit food, and spend two days without seeing a member of the fairer sex, but at the end of the day it’s about meeting people, new and old, and having great conversations about betting.

I’ll see you at Smart Bash 2027.

I read a lot of books about gambling, trading, computing and - to a lesser degree - sports. I thought I’d take this opportunity to share with you a list of such books I’ve been curating over the years. I was careful not to use the phrase “reading list” in the title of this article as that’s a loaded term that includes academic papers and - in this modern world in which we live - Substack and X accounts. I’ll share recommendations in these mediums - especially academic papers - with you in the future.

Links are provided to each book on Amazon (UK) - click the title or the image - but I’m not earning any affiliate fees (I’m not that much of a hustler). If you want to support my work, simply upgrade to a paid subscription.

Get Started Here

The Logic of Sports Betting - Ed Miller and Matthew Davidow

I aspire to write a textbook about sports betting - the kind of book that every new quant might be given on their first day at their first job. A book that explains the mathematics of odds, the landscape of the betting market, how to go about building different statistical models for different sports, and what to do about staking. It’s partly why I started this Substack - the idea being that I could write a series of articles forming the basis of the book’s chapters, a way to tackle the otherwise impossibly daunting task in incremental fashion. The only time I’ve ever sat down and tried to write a piece explaining odds from scratch - really from scratch, not an advanced discourse on the merits and pitfalls of different devigging methods - I turned to “The Logic of Sports Betting”, a stalwart of my virtual bookshelf, to see how Ed Miller and Matthew Davidow did it, and how I might add to what they’d already published. The conclusion: I couldn’t. I had nothing further to add to this excellent primer.

Now, there are some caveats. The book is very much US focused and that means the odds are expressed in US format which can be commended for nothing other than taking an already challenging concept and making it completely incomprehensible. The book doesn’t cover everything an aspiring sports bettor might need to know. Specifically, it doesn’t talk about building statistical models for sports. Arguably these aren’t essential as there are countless successful professional gamblers that don’t originate their own prices. But at some point you are going to need to do statistical analysis of some form even if it’s not to make predictions. Fortunately there are many other books that cover this topic, and you can find some of my recommendations below. But if you’re about to take your very first step on your professional gambling journey, this book is an excellent place to start.

Statistical Sports Models In Excel - Andrew Mack

If you’ve never built a statistical model before, this book is for you. Even if you’re a seasoned modeller, I suspect you’ll still have something to learn from ’s writings. By necessity you’ll learn both about the practical implementation of models and the different types that exist - and which one may be best suited to modelling a given sport. Andrew has a follow up book which is also well worth a read, and he now opines on options trading, both in podcast appearances and book format, the latter being a collaboration with the legendary Euan Sinclair entitled “Retail Options Trading”. That one is also worth a read, at the very least for the second chapter on “The Essence of Edge”, a universal concept equally applicable to betting markets as it is to options trading.

Think Like A Professional Gambler

A Man For All Markets - Ed Thorp

I aspire to be Ed Thorp - a professional gambler who, as the name of his autobiography suggests, has beaten pretty much every market he’s turned his attention to. From blackjack to roulette to horse racing to the biggest casino of all: Wall Street, Thorp is a living embodiment of the idea that markets are markets and value is value and if you can model why a horse is mispriced you can do the same for convertible bonds. This is one of those books that’s worth coming back to again and again for a regular reread as you’ll pick up something new every time. Not nearly as practical as other books I might recommend to beginners but a great source of inspiration, and insight into the mindset needed to beat any kind of market, whatever the asset might be.

On The Edge - Nate Silver

Is there a more recognisable statistician than ? Possibly only that cunt Taleb. Apparently they’re both divisive figures although in Silver’s case I don’t see it: what’s to hate about this cherubic, articulate, endearingly geeky political forecaster cum journalist cum professional poker player? His first book, “The Signal and The Noise” is an equally commendable, accessible introduction to statistical forecasting. I remember buying that book for my boss at the first betting syndicate where I worked so that we might better communicate in the language of statistics.

As for “On The Edge”, I adore the fact that this book exists. It’s like a love letter to professional gamblers - not just sports bettors but also venture capitalists, crypto bros and hedge fund managers. In the parlance of today’s youth, it makes me feel seen. Silver takes us on a fascinating tour of these worlds, highlighting the commonalities in the ways these disparate risk takers think. It’s the central thrust of the book, a useful framing that could help inform your own approach to gambling. Or just read it for the warm fuzzy feeling of knowing that, at the end of the day, we’re basically all the same sharks, we just swim in different waters.

Genuine Alpha

Beyond The Odds - Elihu D. Feustel

This is a deeply practical book, overflowing with valuable advice on a diverse range of topics. Unfortunately, a lot of the latter half is focused on how to approach beating US bookmakers which, as a UK based exchange bettor, is largely irrelevant to me. Nevertheless, it’s made it onto my recommended book list for ’s “Answer Key” concept which is the clearest example I’ve found of a market beating model printed in black and white. It seems to me to be very closely aligned with Empirical Bayes where the prior comes from the market, and the de facto way to be pricing alt lines and derivative markets. Maybe even player props. This alone is worth the price of admission although I’d be shocked if it’s your only takeaway from Feustel’s wealth of experience in printed form.

Precision - C X Wong

For a while there, it looked like “Precision” by C X Wong was permanently out of print. That would have been sad because it’s my go to recommendation for anyone wanting to get into or improve their horse racing modelling. It’s a book dripping with mystique. We’ve got this wonky translation (from Chinese) of a supposedly best selling book from Hong Kong and who the fuck is this C X Wong guy anyway? It’s almost certainly a pseudonym. All of the links - at least in my edition - are dead. In the preface he claims to have hung out with Bill Benter. And yet if you look past all of that and just get stuck into the material it’s clear he knows what he’s talking about. This is a serious book rigorously covering real, mathematical concepts in horse racing, written by someone who - despite the anonymity - all evidence points to having spent years if not decades working for syndicates in Hong Kong.

Aside from the quality of the translation, the chief complaint about the book is that it’s “too mathematical.” Don’t be That Guy. Frictions like these create opportunities. That’s exactly why I’d say this book contains “genuine alpha”. You just need to be comfortable with multidimensional integration to unlock it.

Staking

But How Much Did You Lose? - Dan Abrams

There aren’t many books exclusively focused on the question of staking. Perhaps because a lot of sports betting strategies are capacity constrained so the answer to the question of how much to stake is simply: whatever the Pinnacle limit is. While that’s a nice - and admirable - place to be, in reality we’re all at different stages in our professional gambling journeys with different bankrolls and different risk tolerances, so a more nuanced treatment is appreciated.

Before ’s contribution “But How Much Did You Lose?”, your choices for a book on the Kelly Criterion were limited to the weighty, exceedingly academic treatment given in “The Kelly Capital Growth Investment Criterion: Theory and Practice” or the pop-science “Fortune’s Formula” better suited to a doting parent wanting to understand exactly what it is their professional gambler son does, with no middle ground between these opposite ends of the spectrum.

In contrast, Abram’s book is very much focused on practicalities. There’s a deceptive amount of material here and I found it slow going - not a reflection on Abram’s writing but rather the conflict between human intuition and concepts like expected growth. It’s worth persevering, however, because these are ideas that could have a profound impact on your trading. Read it, understand it and internalise it.

Lessons From Financial Markets

The Laws Of Trading - Agustin Lebron

I followed Agustin Lebron on Twitter - sorry, X - long before he wrote a book - although now that I think about it, it probably was still called Twitter at the time… Ah, fuck it. While not as bombastic as, say, but equally not as much of a larper as, say, Stat Arb, he clearly chats some good shit. So I was thrilled when he released a book, finally breaking free of the 140 character bonds of his original medium. You’ve probably already picked up at least some of his ten “Laws of Trading” (wait, is it eleven?) unless this is literally the first article you’ve ever read on trading but even if you’ve come across them before, you probably haven’t had them explained in such an erudite way. They’re things like: think about who your counterparty is; what is edge? and the importance of technology. It can be incredibly powerful to be reminded of such fundamental concepts, concepts your should know, you obviously know, you’ve been doing this a long time after all, but maybe you’ve forgotten all the gritty details because you’ve been too consumed thinking about prediction markets and AI. It’s a book to be read and read again, to be kept close to your desk and referred back to in moments of reflection. In fact, I’m going to start re-reading it as soon as I hit publish. And I’m excited to do so.

Expected Returns - Antti Ilmanen

Some of the books on this list could span multiple categories. “Expected Returns” by Antti Ilmanen is one such book, being equally well described as “Genuine Alpha”, at least in the more casual interpretation of the term as “something that makes money” rather than the technical definition of “something uncorrelated with market returns.” The book is centred around all types of risk premia - equity, carry, volatility - and Ilmanen explains how these arise due to behavioural biases, institutional constraints and the limits of arbitrage.

There’s two key takeaways for us bettors. Firstly, we - everyone, in fact - should be building their investments around harvesting these risk premia. Secondly, we should consider how the same phenomena that give rise to these risk premia in traditional financial markets might also present opportunities in betting markets. Behavioural biases like punters seeking lottery ticket style returns or preferring to bet on exciting outcomes can give rise to mispricings we could potentially exploit, as can institutional behaviour like hedging lopsided risks although this one is the weaker analogue. Finally, the canonical example of limits to arbitrage in betting markets is the Mayweather-McGregor fight where the huge sums of money wagered actually acted to push the market away from, not towards, efficiency.

How to Build Things

A Philosophy Of Software Design - John Ousterhout

I’ve got a colleague at work that writes disgusting code. Like genuine, absolute dogshit. And AI has only made the problem worse not better. He gets away with it for two reasons. Firstly, he’s one of the nicest guys I know. Honestly, top top bloke. Secondly, he’s a crazy hustler. He gets shit done. It might make you want to blind yourself to take away the pain and it’ll be nigh on impossible to reuse or extend but by god it’ll deliver the goods.

For all you Pythonistas out there, I’d like to share with you a real banger I had to deal with in my latest code review that sums up what I’m talking about:

for _, selection in enumerate(selections):

Actually, it was worse than that because holy fuck he needs to read Chapter 14 of “A Philosophy Of Software Design”, entitled “Choosing Names”:

for _, sel in enumerate(sels):

Obviously what he actually needs to do is read the whole fucking book. As Chapter 3 succinctly puts it: “Working Code Isn’t Enough.” If you think that’s an eyebrow raising statement, this book is especially for you but, really, it’s for anyone that writes software. It’s a classic for a reason.

Designing Data-Intensive Applications - Martin Kleppmann

OK, I have a confession to make. I haven’t actually read this one - although I hasten to add that I do own it. I guess I’m just addicted to buying product. I bought it when I was working on a collaboration on South African horse racing where I had to build an API to provide data for the day’s racecard and that project sadly ended up fizzling out.

It’s a given that at some point in your professional gambling career - if you’re going to have any chance of making it - you’re going to have to build something that could well be described as a data-intensive application. A lot of beginners assume that the only thing you need to make money betting is a good model, and that what makes a model good is the structure and the process used to build it. Contrast a beautifully hyperparameter tuned XGBoost model with a run of the mill linear regression. However, this overlooks two things.

Firstly, the model is actually one of the least important things in the process. A model is useless if you have no way of getting on. You can lose money with a good model and a bad staking plan. You’ll need some level of automation if you want to have any hope of scaling to meaningful amounts of money. And there is a lot more that goes into a successful betting operation than the placing of bets. Keeping detailed records for your accountant, for example.

Secondly, almost everything that makes a model good is the inputs - the data - not the structure. Sure, if your problem is nonlinear it might be a good idea to use a model that can account for that but, as the saying goes: garbage in, garbage out.

So be thoughtful about your data. And read this book. Do as I say, not as I do.

Sports For People Who Don’t Care About Sports

Chasing Points: A Season on The Pro Tennis Circuit - Gregory Howe

I don’t have much interest in sports - despite my income being so heavily reliant on them. Certainly the rules are critical to understand but I don’t have a favourite sports team, I rarely watch live sport, and I have absolutely nothing to contribute to soccer themed office banter. So I read very few books on sport that have zero gambling content but one I have read and adored was Gregory Howe’s “Chasing Points” about his year spent trying to make it as a professional tennis player. It’s a hauntingly romantic portrayal of him playing at the very edges of professional tennis, travelling to far flung places around the globe - places you might imagine would struggle to have running water let alone tennis courts - trying to achieve that single elusive ranking point that forms the bon mot of the title. It reminds me of reading Cormac McCarthy (bear with me) and other authors that write “Westerns” set deep into the 20th century. That’s because the world Howe describes no longer exists. Reforms to the tennis ranking system have made it impossible now for someone like him - a talented amateur - to break into the professional tennis circuit. Read it and weep for what we have lost but also for the uplifting message that it is never too late to follow your dreams and - yes - another story about how systems (in this case: tennis ranking points) can be probed, understood and ultimately beaten.

I started writing an article about it. Not about what I’d done to achieve this record - because I hadn’t done anything - but about it being an inspiration to talk about ways in which I could further improve the flagship strategy, and ways in which I could take one of the strategies I’d shelved and get it over the line and back into production. Both of these parts required their own research. I set about writing and researching then writing some more and researching some more.

Days turned into weeks and then the Cheltenham Festival happened. The strategies printed another weekly profit record. I had to add a note to the start of my draft article explaining that the previous weekly record, the one the article was referring to, actually predated the Cheltenham Festival and had nothing to do with its lashings of dumb money. The article still wasn’t close to being finished, the process mired in inconclusive research findings that were compounded by hastily written code in an attempt to get it out the door.

The following week the whole picture shifted dramatically again. The strategies set two more records: for the worst weekly loss and the biggest single losing market. The latter obviously contributed to the former but even after excluding it, it was still a definitively losing week. This was especially disturbing: I practically hadn’t had a losing week for over five months. The big losing market was down to a disqualification. These have hurt me in the past but I’ve come to view them as the price of doing business. They’re very much evocative of a risk premium.

Compounding the misery were operational issues which meant I wasn’t trading a large proportion of the strategy portfolio for a period of three days. The week of Cheltenham, Betfair rolled out their TPD1 coverage to all RMG owned race courses in the UK. After observing discussions about the dubious data quality associated with these race courses on the betcode Slack, I decided I had to deploy a code change to reduce stakes on these races. I was reluctant to do so as the flumine trading instance hadn’t been restarted in months, and there were a couple of significant, breaking changes that had been made to the flumine codebase in the interim. But it felt better to get the change deployed even given these circumstances than to trade the RMG races at full stakes right off the bat.

Unfortunately, due to inadequate testing ahead of deployment, my code change resulted in an uncaught exception which caused the flumine process to exit. I only noticed three days later after thinking that the volumes traded by the strategies looked suspiciously low - as opposed to receiving an explicit alert after the process exited. In a further embarrassingly amateur twist, my poorly set up logging infrastructure meant the relevant log files had already rotated with valuable diagnostic data lost forever.

The strategies then started losing money every single day. This is not behaviour consistent with a portfolio that hadn’t had a losing week for five months. The obvious smoking gun was the initially botched then rectified deployment. Did my own code change introduce another, more subtle bug than the one which caused the uncaught exception? Or was the change exogenous? Had whoever been on the other side of my record Cheltenham week decide to finally pack it in? Had my edge evaporated overnight? It smacks of hubris but the paranoia had started to set in.

The thing that really disturbed me was not that the realised return had fallen off a cliff - but that the expected value had also flatlined. All other metrics looked fine: number of betting signals, fill rates, average liabilities - I couldn’t find a single way in which the behaviour of the system had changed - aside from the fact it no longer seemed capable of making any money. I lined up a battery of backtests - to compare simulated and live bets on the same markets, and to search for historic runs of cumulative expected value that looked as bad or worse than what I was currently experiencing.

The pain was relentless. Going to run the backtests, I was assaulted by warning after warning that the required data files not did exist in my S3 bucket. What the hell was going on? It turned out my data scraping process for the Betfair race stream had broken two months ago without me realising. Last year - when I was actually actively developing these strategies - I would frequently run backtests comparing live and simulated results, a key step in the long road from idea to production, and one that would have immediately alerted me if I was missing any essential data like these scraped Betfair race stream files. But I’d never automated the process and now I was paying the price.

Thankfully, I’ve cultivated a network of other Betfair-first professional gamblers and I was able to reach out to one of these to help me in backfilling the missing data, for which I am greatly appreciative. The backtests provided some measure of reassurance. Those run on the same recent races showed results bang in line with what happened in live trading, and the ones looking for poor historic runs showed me that poor runs of expected value like what I was currently experiencing were not a rare occurrence. Now the question was: why? How could the behaviour of the system - the nature of the bets it was placing - look unchanged yet for them to be generating virtually no expected value? That had to be something exogenous, surely? Something in the market had to have changed.

The answer turned out to be ludicrously simple. One final embarrassing admission, a conflux of total ignorance of the sport I was trading and the fallibility of human memory. I know enough to know that jumps racing is very different to flat, that Cheltenham Festival is a jumps event, and to speculate that there might be something a bit different about the horse racing calendar following such a premier event. I checked the proportion of jumps races since everything went to shit and noted it hadn’t been as high as it was for… about five months. I checked my backtest results and thought: “huh, I’m not sure this strategy actually makes money on jumps races”. I split my expected value graphs by jumps and flat: flat looked fine, jumps was in the toilet. Something was tickling the back of my brain. I looked up at the whiteboard in my study, where I’d made some notes about this strategy last year, what felt like a lifetime ago, subsequently ignored as my professional world became consumed by prediction markets, and read the mocking words, carefully penned in red ink:

“Doesn’t seem to work for jumps”

I hope you’ve appreciated my openness in this telling. On the one hand, it’s painful to reflect on the inadequacy of some of my processes like alerting, logging, and testing. On the other hand, I pride myself on taking a pragmatic approach to my trading. Up until now, these details haven’t mattered. I’ve taken a beating but it’s not remotely existential. In a way, discovering that the botched deployment had absolutely nothing to do with the downturn in fortunes has given me confidence to start making some of these long needed changes. I’d been putting off making any changes that required a restart of the flumine process, too afraid to even breathe on the money printing machine that was delivering profits week in week out. Now that the choice to restart the process has been taken away from me, and that the strategies have most definitely survived the restart, I’m ready to tighten up my logging process and implement some much needed alerts.

Those are definitely some of the opportunities that the title of this article refers to. But over and above that, remembering the poor performance on jumps racing opens up some more exciting angles to explore: have strategy parameters been made to awkwardly fit both jumps and flat racing when flat racing could perform even better if it wasn’t held back in this way? And more importantly: can jumps racing be turned from a drag on the portfolio to a profitable contributor in its own right? I’m excited to find out.

Now you know why I haven’t managed to publish anything in the past month. However, I have an exciting pipeline for April which I hope makes up for this:

1. The release of the Algorithmic Sports Betting GitHub repository containing data and code for all published research articles.

2. A recording of my recent presentation at the London School of Economics on 26th March 2026 entitled “Revisiting: RSS Wimbledon 2022 Prediction Competition - A Professional Gambler’s Approach”.

3. A research article quantifying the advantage conferred by winning the coin toss in a tennis match.

I look forward to sharing these with you in due course.

This is one of those big, beautiful edges that can be understood by a five year old and vibe coded by anyone with a ChatGPT subscription. But does it actually exist? I spent some time digging into the paper. I found tantalising backtest results which - in their academic ignorance - the authors almost immediately dismissed. As I dug further, the gap between academic understanding and practitioner insight seemed to widen, with questionable handling of data quality issues, ignorance of dead heats, and a truly pernicious lookahead bias.

I set about replicating the analysis, bringing it up to date with the additional seven years of data now available since it had been published. At first blush, the bias appears to have weakened over time - or perhaps it never existed in the first place. With the limited data available and such a small effect size to be measured, it becomes challenging to tease apart signal, noise, and regime changes. With an edge this small, execution details are critical and, having addressed the shortcomings of the original analysis, my backtest results show promising hints of profitability. Let’s take a look.

The Data

Merz et al. used a freely available dataset of Betfair Starting Prices (BSP) published by Betfair in order to carry out their analysis. The BSP of a horse can be considered to be its fair (“true”) odds of winning the race before it starts, analogous to the “closing line” in US sports.

The way in which the BSP is calculated is rather esoteric but it is essentially the result of an auction where backers and layers commit funds and the degree of imbalance between the backers’ stakes and the layers’ liabilities determines the odds. It is in many ways the same as the opening auction on a traditional financial exchange.

Interested readers may further plumb the depths of the BSP calculation rabbit hole at their leisure but for our purposes the key takeaway is that this published BSP dataset facilitates a wide range of analyses and can form the starting point for strategy development if you do not have expensive Betfair price stream data at your disposal.

The paper focused on races from March 2008 to May 2018 across the UK, Ireland, the US, South Africa and Australia. Since 2023, Betfair also publish data for races in France and the UAE. However, not every race on Betfair features in this BSP dataset. For example, New Zealand races are excluded for reasons unknown. In addition to horse racing, data is also available for greyhounds.

As well as the BSP for each horse, the dataset contains key information like the horse’s name, a win-loss flag, a weighted average price (WAP) for all bets struck before the race started, the same for all bets placed in running, minimum and maximum odds matched, and volume information.

I tried very hard to reproduce the exact dataset described in the paper. Detailed figures are provided for the total number of races and total number of horses but specific dates are not given beyond that the dataset covers March 2008 to May 2018. I even experimented with every permutation of a starting date in March 2008 and an end date in May 2018 without success. It’s possible that subsequent to the analysis, Betfair updated the available data for this period. At the very least, I do get good agreement with some of the summary statistics mentioned, such as there being, on average, 9.2 runners per race.

In the end, they cite 4,066,455 horses across 443,850 races, prior to doing any data cleaning. My numbers are 4,117,622 and 447,789 respectively.

While this BSP data is genuinely valuable, it does have some quality issues. The paper states the following:

In approximately 900 races, there were multiple winners. Our results are insensitive to the inclusion of those races. A value for the WAP or SP is deemed erroneous if the price is below 1.01 or above 1000, which corresponds to the odds range given by Betfair. Our results do not change if we exclude only the missing values for the WAP or SP instead of the whole race.

The problem is that neither the multiple winners nor prices outside the standard Betfair odds ladder are actually errors, and I discovered much more egregious problems that the paper fails to address:

1. Data from some markets other than the race winner have snuck in, including but not limited to: some place, forecast and reverse forecast markets

2. Some races only have a single runner

3. Some races have no winner

4. Some races have zero total volume matched across all runners, which is not credible

It seems impossible to avoid the conclusion that Merz et al. overlooked some if not all of these issues as their data quality filters only reduce the number of horses in their dataset from 4,066,455 to 3,903,604 whereas mine reduce from 4,117,622 to 3,274,997.

Regarding multiple winners, horse races can and do have multiple winners according to deadheat rules. Two and, indeed, more horses can tie for first place. In the case of a two way tie, for example, backing one of the winners is treated as a half win and a half loss. It’s not at all clear from the exposition whether Merz et al. appreciated these mechanics and simply chose not to have to deal with the complication, or lacked the domain expertise to understand what they were seeing in the data.

As for BSPs outside the limits of the standard Betfair price ladder, I would allow that a BSP of less than 1 is nonsensical but, given the way BSP is calculated, the prices do not conform to the standard ladder at all and can be further distorted by the impact of non-runners. Under some circumstances, some horses will attract no BSP bets and therefore no BSP can be calculated. Conceptually that horse has zero probability of winning and the BSP should be infinity.

It’s a bugbear of mine when researchers say things like “Our results are insensitive to the inclusion of those races.” One man’s insensitivity is another man’s edge. I get that scope needs to be tightly controlled - especially as I work more and more on my own Substack writings - and that not every possible research avenue can be fully explored before publishing results. However, I’d much rather see some specifics on how the numbers change as deadheats are and are not included and make up my own mind as to whether it makes a meaningful difference.

As well as preparing a dataset that resembles the original as closely as I could manage, I also expanded it with the seven years of data that has accumulated since the paper was originally published. As with all of my published research, paid subscribers can get access to the supporting data and code. While this BSP data is publicly available, it’s not very well organised and it’s quite tedious to download large chunks without running afoul of rate limits. As mentioned, there are also some data quality issues which I’ve cleaned up for you. Please be on the lookout over the next week or so for details on a dedicated Algorithmic Sports Betting GitHub repo.

Testing For the Effect

The paper sensibly breaks the analysis down into two key stages:

1. Testing that the effect exists

2. Finding adequate evidence that it does, testing whether the size of the effect is economically exploitable

This is a sound way to approach any kind of strategy development. Don’t waste time implementing and backtesting a strategy without first building a convincing case for doing so. You want to be failing fast in your search for the most promising edges.

Testing for the effect is very simple. They perform a logistic regression where the target is the win/loss flag for the horse and the predictors are:

1. An intercept term

2. The inverse BSP for the horse

3. A binary flag as to whether the horse has a fast-sounding name

In other words, they test if a horse having a fast-sounding name has any residual predictive power over and above the probability implied by the BSP.

Using the raw inverse BSP rather than removing the overround (devigging in US parlance) seems suboptimal to me as it requires the logistic regression to learn both how to calibrate these unfair probabilities as well as fit the residual fast-sounding name effect. The authors in fact carry out a few different permutations of the analysis and in one of these they replace the inverse BSP with the implied probability calculated by removing the overround using the “standard” method of dividing by the booksum. I have many thoughts on the best way to carry out overround removal that will have to wait for an upcoming article but for now let me state that I am a big fan of Shin’s method which can better account for phenomena like favourite-longshot bias.

The logistic regression presented in the paper has a fitted value of -0.002 for the fast-sounding name effect, significant at the 5% level. The negative sign is consistent with the hypothesis that these horses win less often than their odds imply. It’s worth highlighting that throughout the paper, Merz et al. analyse both the BSPs and the WAPs. The latter seems irrelevant to me as:

1. You cannot execute a single bet at this price

2. It includes trades far away from the off and is therefore naturally less efficient than the BSP

In fact, including it in the analysis weakens the case Merz et al. seek to make, as follows. In total, 12 regressions are carried out: 6 fitting the WAP and 6 fitting the BSP, using both different families (logit, probit, OLS) and different predictors. This is a classic case of the multiple testing problem. The more hypothesis tests that are carried out, the more likely it is to observe statistically significant results by chance alone. The evidence for the fast-sounding name effect was already weak, being only significant at the 5% and not 1% level, and the sheer number of permutations of the analysis carried out makes it even less likely the effect is genuine and not simply random noise.

I’ll now present my own results testing for the presence of the fast-sounding name effect. I’ll start by attempting to reproduce the paper’s analysis as closely as possible. I’ll use the same definition of fast-sounding name as used in the paper although, as previously mentioned, my dataset does not include exactly the same set of races. I’ve also removed some junk data that most likely polluted the paper’s results. I suspect that any discrepancy between my findings and theirs will be down to this.

I’ll then show updated results for March 2008 to February 2026 and explore whether and how the effect has changed over time.

Replicating the Original Analysis

                           Logit Regression Results                           
==============================================================================
Dep. Variable:               win_lose   No. Observations:              3274997
Model:                          Logit   Df Residuals:                  3274994
Method:                           MLE   Df Model:                            2
Date:                Sun, 22 Feb 2026   Pseudo R-squ.:                  0.1580
Time:                        15:16:50   Log-Likelihood:            -9.4593e+05
converged:                       True   LL-Null:                   -1.1234e+06
Covariance Type:              cluster   LLR p-value:                     0.000
================================================================================
                   coef    std err          z      P>|z|      [0.025      0.975]
--------------------------------------------------------------------------------
const           -3.1987      0.003  -1224.376      0.000      -3.204      -3.194
inverse_bsp      7.3172      0.014    510.067      0.000       7.289       7.345
is_good_name    -0.0192      0.008     -2.389      0.017      -0.035      -0.003
================================================================================

Conditional on BSP, horses with fast-sounding names have a 0.0019 lower probability of winning, almost identical to the 0.002 figure reported in the paper.

All Data to Date

                           Logit Regression Results                           
==============================================================================
Dep. Variable:               win_lose   No. Observations:              8098465
Model:                          Logit   Df Residuals:                  8098462
Method:                           MLE   Df Model:                            2
Date:                Sun, 22 Feb 2026   Pseudo R-squ.:                  0.1611
Time:                        14:55:13   Log-Likelihood:            -2.3636e+06
converged:                       True   LL-Null:                   -2.8175e+06
Covariance Type:              cluster   LLR p-value:                     0.000
================================================================================
                   coef    std err          z      P>|z|      [0.025      0.975]
--------------------------------------------------------------------------------
const           -3.1934      0.002  -1909.639      0.000      -3.197      -3.190
inverse_bsp      7.2884      0.009    808.153      0.000       7.271       7.306
is_good_name    -0.0073      0.005     -1.423      0.155      -0.017       0.003
================================================================================

Expanding the data to include the 7 additional years of horse racing since the paper was published shows a weaker effect both in size and significance. Given the convincing result for data up to May 2018, this implies that the effect has been much less pronounced since then. Indeed, carrying out regressions on a year-by-year basis shows that the effect has perhaps disappeared as early as 2018 and more convincingly by 2021. However, take note of the standard error bars which remain quite wide due to limited sample sizes when slicing the data in this way.

Backtesting

Merz et al. carry out a backtest as follows. They start by considering backing every horse with a fast-sounding name, risking 1 unit per bet. Assuming they get filled at the BSP, this yields an average return of -0.0289 units per bet before commission. It’s unclear at this stage how much this could be explained by the BSP being overround, and how much it can be attributed to these horses being potentially overbet.

So far, the setup is largely uncontroversial. In reality, there would be some slippage such that the filled prices should be a bit worse than the historically observed BSP - i.e. the BSP would move on the basis of the bet - but if we can’t make the strategy work before slippage and fees, it’s certainly not going to make money afterwards.

The next part is where things start to get really hairy. To profitably trade this effect, we’d want to lay horses with fast-sounding names, not back them. To calculate the lay profit, Merz et al. simply take the negative of the back profit. In other words, they assume an average return of 0.0289 units per bet, before commission. They’re continuing to assume that the backer is risking 1 unit per bet but as the layer they stand to win 1 unit per bet and risk an amount that is dependent on the odds.

If I reproduce the equity curve for this strategy using my data then it looks pretty good: up and to the right albeit with a few precipitous drawdowns and some plateauing after 2021. That’s consistent with the findings of the previous section that the effect seems to have diminished after this point. Also note that the paper reports a total return of about 5,000 units which is roughly where my curve sits at the end of April 2018.

Far worse than ignoring slippage, this backtest suffers from a devastating lookahead bias. When you lay a bet, your liability is:

Where the odds are expressed in decimal format and stake refers to the backer’s stake. When placing a BSP bet you don’t specify the odds - you specify the liability. You commit to how much you are prepared to lose. But Merz et al. fixed the backer’s stake - the amount a layer stands to win - to 1 unit. In order to calculate the liability that this implies, it is necessary to know the odds. But with BSP, you don’t know what the odds are until the auction has taken place. Conceptually, the paper took advantage of knowing the true odds of the horse before the race started.

To further illustrate the benefit of the lookahead bias, here’s the equity curve when laying horses without fast-sounding names:

In other words, laying all horses all of the time is a profitable strategy. Having said that, there is some evidence of excess returns attributable to horses with fast-sounding names as these yield an average per bet return of 0.0218 units while other horses return 0.00909 units on average.

For a realistic backtest, rather than laying-to-win 1 unit per bet, you have to risk 1 unit per bet. The equity curve for that looks very different:

Read more

This is my first paid article and represents the kind of deeply researched, actionable insight I intend to put out on a monthly basis. In this instance, only a small portion of the article will be paywalled, so readers can get a good idea if the material would offer value to them. Paid subscribers may also contact me for a copy of the derived, pre-processed dataset and code underlying the analysis and I expect this to be a feature of future research articles as well.

Methods

The analysis was carried out on Betfair price stream data recorded between March 2022 and January 2026 for all UK horse racing win markets. I combined it with data scraped from the undocumented Race Card endpoint which provides useful metadata including, for the purposes of this analysis, the race course, the type of race (jumps, hurdles, flat), the race class, whether the race is a handicap, and the race distance.

I used the betfairutil Python package (written by yours truly) to efficiently extract the total amount traded up until the race started, and the total amount traded by the time the race ended for all races. I then built two Generalised Additive Mixed Models (GAMMs), one to regress the pre-off volume traded, and the other to regress the in-play volume traded, taken as the difference between the total volume traded and the pre-off volume traded.

If you are unfamiliar with them, Generalised Additive Models are linear models where the response variable is modelled as a linear combination of functions of the covariates. So, for example, rather than assuming the race distance linearly relates to the in-play volume traded, we can fit a flexible spline that can capture more complicated relationships like a plateauing effect.

The Mixed in Generalised Additive Mixed Models refers to the separation of the factors into those which are “fixed” and those which are “random”. Without getting caught up in the statistical weeds, the race course is an ideal candidate for modelling as a random-effect as the dimensionality is large and the sample sizes are heterogenous. Conceptually the race courses observed are “random” samples from the population of all race courses but while statistically correct, this framing is unintuitive as we do, in fact, observe every race course. A better way to think about it is in terms of the number of parameters in the model: rather than modelling each race course independently and fitting a separate parameter to each, we model how much race courses differ from each other, requiring only a single variance parameter to be fit.

The goal with these regressions is to isolate uninteresting factors contributing to the volume traded - for example, it’s obvious that higher quality races will attract more money - and attribute the residual volume to more interesting effects like those associated with exploitable human biases.

Factors

This is the full list of factors included in the two models, along with an explanation for each. I’ve grouped the factors into some broad categories.

Race Quality

• Race type. This can take the value Jump, Flat, Hurdle, and National Hunt Flat

• Race class. A value from 1 to 6, where 1 is the highest quality of race and 6 is the lowest

• Handicap/Non-handicap. A boolean variable indicating whether the race is a handicap race

• Number of Runners. How many horses actually participated in the race

Race Timing

• Hour of Day. After accounting for race quality, whether there is any residual effect from races scheduled at more convenient times of day, i.e. the evening

• Bank Holiday. A boolean flag indicating if the race fell on a public holiday in the UK

• Day of Week. Similar to hour of day, after accounting for race quality, whether there is any residual effect from races scheduled on particular days, especially weekends compared to weekdays

• Year. Capturing the long term trend

• Day of Year. Capturing seasonality within the year. Note that the impact of big festivals like Cheltenham are captured by other factors, as is the calendar effect of jumps season so the expectation is that the residual seasonality is negligible

Behavioural Effects

• Last Race of the Day. It’s well known that the end of gambling sessions attracts more money from loss-chasing punters placing “double or nothing” bets in the hopes of ending up flat after experiencing losses. It’s plausible that in a horse racing context this would manifest as increased volumes bet on the last race of the day

• Days to Pay Day. Immediately following pay day there’s expected to be more uninformed money flowing into the markets. 0 represents pay day itself and high values of this factor correspond to the days just after pay day

• Number of Races Today. More races held on the same day dilutes available funds and, from a human perspective, attention

Random Effects

• Race Course x Month. This is the only interaction term in the model and captures when big meetings like Cheltenham Festival are held throughout the year

• Race Course. By including race course as a separate factor as well as in the interaction term with month, we can determine whether, for example, more mundane meetings at Cheltenham attract more money compared to other venues

In-play Effects

• TPD Coverage. A boolean flag indicating whether Total Performance Data (TPD) is covering this race. In case you are unaware, TPD is a data provider that uses GPS trackers placed in the horses’ saddles to provide real-time positioning, velocity, and stride frequency data as the race takes place. I included this factor in both the pre-off and in-play model, as I wanted to see whether the potential to use this data in-play leads to traders taking on bigger positions before the race starts

• Race Distance. Very simply, the longer the race distance, the longer it takes to run and the more time there is to bet during the race. This factor was only included in the in-play model

It’s worth highlighting Prize Money as an example of a factor which didn’t make it into the model. An advanced reader kindly highlighted that this is a potentially much finer-grained measure of race quality than Race Class. I’ve added it to the dataset available to paid subscribers and I encourage them to explore the impact of including it in the model.

Results

Big picture, the year-on-year changes in volumes are grim. After a relatively modest decline of around 10% in 2023, these fell further in 2024 to 75% of 2022 volumes. Remember that these figures are after controlling for race quality in various ways which means the decline cannot be explained by, for example, fewer high profile race meetings being held or a reduction in average field size.

Instead, I suspect what we are seeing is at least partly down to affordability checks limiting inflows of soft money. These officially came into effect at the end of August 2024 but may well have been pre-empted by Betfair earlier in the year in the hopes of getting ahead of the regulator. Pre-off volumes stayed relatively steady in 2025, lending weight to the affordability checks introduction thesis as does the fact that the theoretically sharper in-play market simultaneously experienced a further decline in volume.

Although we are less than a month into 2026, the initial signs are concerning with pre-off and in-play now below 70% and 60% respectively of 2022 volumes. This is rampant speculation on my part but I do wonder if in-play in particular is suffering from automated bettors switching their focus to prediction markets.

Including the number of races held on the same day as a factor in the model facilitates quantification of attention fatigue: the decline in volume traded attributable to too many races dividing punters’ attention and competing for their available funds. The above plot illustrates the kind of informative functional summaries that drop out of a GAMM. Per race pre-off volumes are on average 40% lower on days with over fifty races compared to those days with at most two race meetings. The pre-off curve is significantly steeper than the in-play one meaning pre-off volumes are affected to a much greater degree by how congested the day’s schedule is. Nevertheless, there is a kink in the in-play curve at forty races suggesting this factor is not just about the limits of human focus but also finite bankrolls.

Modelling the number of days until payday captures the effect of jumps in punters’ available funds. Payday itself is considered 0 days to payday and races held on this day experience 4% less volume matched than the average. Given payday is, by definition, a workday, only the most compulsive gamblers - or the keenest fans of the sport - would be able to find the time to deploy significant amounts of their newly acquired capital on this day itself. Instead, we should see an increase in volumes matched immediately following payday which we do. Days to payday peaks just after payday and pre-off volumes are up over 8% compared to the average race on these days. In-play volume is also up but not quite to the same extent.

The model produces sensible results for the race course x month interaction effect. Far and away the standout combination is Cheltenham in March corresponding to the Cheltenham Festival, attracting over 250% more volume than the average. This is followed by Ascot in June (Royal Ascot), Aintree in April (Grand National), Goodwood in July (Glorious Goodwood), York in August (Ebor Festival) and Epsom in June (Derby Festival).

Even after accounting for the big meetings with the race course x month proxy, the big race courses still attract outsized volumes. Cheltenham is, again, the standout here with almost 80% more volume compared to the average, followed by the usual suspects of Ascot, York and Epsom. The race course with the largest negative effect is Cartmel with a bit more than 10% less volume than the average. Many of the race courses with the lightest schedules feature in the bottom 10 by volume. Note that the model effects combine multiplicatively so a Cheltenham Festival race has the uplift from both the race course x month interaction effect and the individual race course effect. All factors are considered relative to the model average, which is a reference point holding all other factors fixed.

Some of the most useful insights come from an interpretation of the model’s fitted parametric effects.

Read more

I never really knew for sure why the original tennis strategy I developed at the start of my career made money. The premise was simple: assume the market is correct at the start of the match but inefficient when it comes to reacting to what’s happening on court. This captured two foundational principles of trading:

1. Use the market to tell you the true price wherever possible. Another example would be using a leading exchange as the source of truth in cross-exchange arbitrage

2. Target periods of market inefficiency. Another example would be betting into softer early horse racing markets rather than the cut-throat seconds immediately before the off

There’s also an implicit third principle as well: target periods of high market liquidity. This is important for being able to scale up a strategy and is contradictory with the idea of targetting inefficiency. Theoretically, the more weight of money behind the market, the more efficient it should be. But in tennis, the pre-match market is pretty dead and liquidity only really opens up once the match starts.

It’s easy to build a case for why the market should be efficient before the match starts and then turn inefficient in-play: before the match starts, human traders have hours if not days to consider soft factors like injuries, motivation, distance travelled, fatigue, home advantage, playing styles, handedness, court speed, and the weather, to name but a few. Some of these things may be hard to measure, or hard to quantify the impact of, or both, giving the human an edge over the machine.

Once the match starts, however, decisions must be made within seconds if not fractions of a second. Soft factors still exist, such as how well a player is moving around the court or other telltale signs of injuries, or manifestations of frustration such as arguing with the umpire. Yet while it may be hard to calculate exactly how much some turmoil in a player’s personal life will affect their chances of winning their upcoming match, once points are being played we are being fed, in real time, bit by bit, information on how well a player is performing: how many of those points they are winning.

In Bayesian parlance, what we have is a prior - the market-implied probabilities at the start of the match - observations - points won along with additional context like breakpoints saved, second serve percentages, number of aces and so on - and a posterior - the prior adjusted based on the observations to give an updated probability. In that co-working space in Dalston, over fifteen years ago, overlooking the canal and with the sun streaming through the windows, I probably sat there and thought that the reason the strategy made money was because the way I calculated the posterior was better than the market. These days, I’m not so sure.

Looking back on it now, I suspect the real edge was simply the automation. I was taking a sensible model and hooking it up with a ropey scores feed with a known seven-second delay intended only for use on websites then gluing them together with proprietary software that could automatically trade every tennis match on Betfair in real time. I thought I was an unbeatable statistics PhD but in reality I was just a plumber.

Here’s some evidence that my model was actually doing something. As I previously wrote about, having “cracked” tennis, that first syndicate of mine moved on to trying to beat association football. I thought I’d found the formula for success so went with the same approach: assume the market is correct before kick-off, then becomes inefficient during the match.

The problem with football compared to tennis is it is much less information-dense. Any given tennis match typically has over 100 points in it, and each of those points is a little update on how each player is performing. In contrast, there are some football matches where you’ll be lucky if there’s a single goal in it.

The natural next step is to think not in terms of goals but in terms of shots. The thing is that shots are not that much more frequent than goals, and shots don’t necessarily matter so much as shots on target. No doubt at this point you’re rolling your eyes and wondering why I haven’t mentioned expected goals (xG) yet but recall that I was working on this stuff in 2011 and the concept of xG was not widely disseminated at that time.

So we were trying to work with shots and shots were not something that was available in the kind of budget data feeds we were using. Instead, we hired a watcher to input data for us. At the same time as having uncertainties about how well the model might work with a limited flow of information, and spending the insane amount of money on compute required to run the model which depended on Monte Carlo simulations, we were now having to worry about how accurate and timely our watcher’s data input was. I still don’t know which of those did us in but it was probably a combination of all three.

In between sports betting jobs I interviewed with a couple of hedge funds. It was interesting to see that my experience building automated betting strategies for Betfair counted for the square root of fuck all. I remember being quizzed about JVM customisations to avoid garbage collection during critical paths and when I tried to explain that I hadn’t bothered with such local optimisations because they’d be swamped by network jitter due to Betfair actively discouraging colocation their eyes started glazing over. The concept of Betfair’s in-running bet delay was equally incomprehensible to them which seems weird in hindsight given speed bumps also feature in traditional financial markets.

I’d forgotten that one of the reasons I took the job with RISQ in the first place was having met one of my prospective colleagues a few months before when he was out on his own trying to pitch a football betting strategy he was looking for investment for. He seemed like and was, indeed, a nice guy and I enjoyed working with him. One thing we bonded over was a side hustle smashing the shit out of Football Index.

RISQ did a little bit of what I guess was venture capital although it was a part of the business I never really got involved in, aside from being in the room for one or two pitches. I don’t think they invested in Football Index but they certainly heard about it early, like soft launch stage, when there were fat VC-backed deposit bonuses available in order to try to attract a critical mass of customers.

The Football Index product was so brain-dead it’s kind of hard to explain but I’ll give a brief overview and you can go watch a YouTube video or something if you feel like you need more details. You would buy shares in association football players and then these shares would pay out dividends in two respects. Firstly, in a fantasy sports kind of way, players would score points based on their on-pitch performances and then the top scoring players would yield a dividend each day. Secondly, online news articles were tracked for mentions of players and the most talked-about player each day would also yield a “media buzz” dividend.

So far, probably doesn’t sound too terrible. The first issue was with the pricing. The prices were entirely artificial. There was no limit order book where one could quote their buy-and-sell price for a player. There was just a market price which moved artificially according to very specific rules based on how many shares of that player were bought or sold. Something along the lines of: if 100 shares were bought, the price would go up 1 p and vice versa. The second issue is that the shares also never (really) expired. I recall there being something in the terms and conditions that they lasted something like 3 years, so that Football Index could satisfy the Gambling Commission that their customers were placing bets and not buying some new piece of shit financial instrument they’d concocted while recovering from poorly carried out brain surgery.

If you’re familiar with Football Index and this description doesn’t sound anything like what you remember then bear in mind that the product was overhauled at least a couple of times before the literal Ponzi scheme collapsed so maybe we just got involved at different times.

I remember making a model but at the end of the day that was completely superfluous as the strategy boiled down to this:

1. Make a large enough deposit to get the maximum bonus

2. Put all of your money into Wayne Rooney

3. Watch as every single day without fail the UK tabloids published story after story about Wayne Rooney, earning you the “media buzz” dividend

In effect, you got a perpetual bond paying you about 3% a day. The only reason I’m not richer than Elon Musk right now is the pricing mechanism meant that Rooney quickly became too expensive. But there was definitely an edge in realising that the lifetime of the shares was too generous and that everyone else was treating this like it was just fantasy football and buying some teenager called Anthony Martial.

At soirees when I tell people I’m a professional gambler people usually assume I’m a poker player and picture me in a smoky poker room risking it all on another throw of the dice (they don’t understand gambling). In reality, there’s about as much glamour to it as any other desk job. But there was one time, one delicious, never to be repeated time, when one of my betting partners handed me a literal envelope full of cash, freshly withdrawn from the Natwest down the road, in those halcyon days before the teller would ask what you needed your money for, and that time which, recalling it now, brings a smile to my face and a replete sense that I have truly fulfilled my identity as a professional gambler, that time was when we settled up after smashing the shit out of Football Index.

I’ve been a professional gambler for over 15 years. You can read all about it in my first post. However, it’s only since the start of 2025 that I’ve been developing and running strategies that are completely my own. Sure, I’ve done plenty of risking my own money, but that has always been as part of a syndicate where the profit, workload and, most importantly of all, the emotional turmoil, were shared.

The thing that held me back from going solo was a lack of imagination. I think it’s important to distinguish between imagination and creativity, the former being the ability to conceive of things that don’t exist while the latter is concerned with bringing those things into reality. I consider myself pretty creative. I like to style myself a “full-stack quant” by which I mean capable of contributing to all stages of strategy development from data sourcing and cleaning to modelling to implementation to deployment. My preferred way of working, and how it tends to work with the syndicates I’m part of and the other collaborations I’ve sought out, is that someone else comes up with the ideas and I run with them.

Towards the end of 2024, however, I had come to a, for me, world-shifting realisation: I could outsource idea-generation without getting into a profit-sharing arrangement. There are ideas everywhere. In books, in academic journals, on fintwit, in the pub with your mates. I don’t mean complete recipes for profitable trading strategies, I mean ideas like: momentum is a phenomenon of markets that has existed for at least three thousand years, first showing up in the price of grain in ancient Babylon, and if it worked for them and it works for shitcoin traders then maybe it works for betting horses on Betfair and oh there’s this concept of steamers and drifters and hmm…

There were two big influences that helped me come to this realisation. First was participating in Robot Wealth’s Trade Like A Quant course. and put out some great material and while the application areas tend to be tradfi or crypto, the lessons are universal to all markets.

The second big influence was the betcode Slack group, a community of like-minded algorithmic sports bettors that has grown up around users of the betfairlightweight and flumine software packages. I’ve been involved in the group for the best part of a decade now and in 2024 I’d started organising monthly meetups. We’d run many events before, including one at Betfair HQ involving a “fireside chat” with the legendary Matthew Trenhaile, but these tended to happen on more of a quarterly basis. I was craving the equivalent of after-work drinks for people like me who work for themselves. We experimented with the cadence and ultimately settled on about once a month.

We’ve had some great discussions in the two years since we started. Everything from the latest book we’re all reading (Dan Abram’s But How Much Did You Lose? being particularly fertile ground) to how much everyone lost fading Trump in 2024, to how much everyone won betting on Biden in 2020 after Betfair left the market open for a month after he’d already won, to the correct way to calculate ROI for lay bets, to the rise of prediction markets in the US, to whether Smart Bash was worth attending (terrible name but in short: yes), to name but a few. I often found myself literally frantically making notes on my iPhone as the conversation swirled around me, eager to follow up later on some nugget of information.

If you feel your sphincter puckering up at that last sentence, imagining me stealing everyone’s trading strategies and laughing all the way to the bank, that’s not at all the case. Sure, I might have discovered that A— makes his money using machine learning to bet horse racing, but I already knew that you can make literally billions of dollars doing that and still haven’t been persuaded that it’s the right approach for me. I find people are way, way too secretive in this game. Details matter. You making me realise I’ve been calculating expected value wrong is not going to cause your profit to evaporate. And your specific strategies just aren’t that interesting to me. But actually yeah, maybe I do want to check how my strategies perform on payday, thanks for that general insight into behavioural economics. I’ll buy you a pint.

While we’re talking about the betcode community, particular shout-out needs to go to Paul who (as well as being one of the loveliest people I know), in December 2024, gave birth to the “Advent of Alpha”, twenty-five hastily written (by his own admission) articles on potential trading strategies. This is the perfect illustration of what I mean about ideas being everywhere. All of these are genuine, honest-to-God ways that you could make money betting. If you’re capable of turning an idea into reality.

This has all been a rather long-winded way of saying that by the end of 2024 I had compiled a list of concrete strategy ideas that I was ready to pursue going into 2025. We’d had an inaugural betcode Christmas party at Rules (London’s oldest restaurant, dontcha know) at which I’d boldly announced that my goal for 2025 was to make at least as much money from these new - at that point non-existent - trading strategies as I made from my existing syndicates. It was a completely ludicrous goal but it gave me a benchmark to compare against. I also publicly committed myself to providing weekly updates on my progress.

This article represents a culmination of that committent. Almost every week of 2025, generally but not always on a Sunday, I would create a post on the betcode Slack with various graphs and a discussion of how the strategy development was going. But Slack is far from the best medium for sharing and - more importantly - preserving these things and that’s why we now find ourselves on this Substack.

The Strategies

As a man who has spent about forty years being fascinated by computers, I’ve only ever publicly referred to the strategies using a numeric code (zero-indexed, naturally). Internally, they have flashy names so, for example, the strategy previously only known to readers as “Strategy3” I can now reveal is actually called QueueJumper. The problem with publicly using the internal names is that they’re also pretty literal and, for most of them, that’s more alpha than I’m willing to leak. QueueJumper is all about getting good queue position and hopefully the benefit of that is so bloody obvious that in this case I’m not actually giving anything away by telling you its name. It was also only in production for thirteen days and didn’t make any money.

For new readers, I probably need to rewind a bit. All of these strategies are betting on Betfair. They’re largely betting on horse racing with a bit of greyhounds and some soccer although that last one is currently shelved. There’s a mix of pre-event and in-running but the best stuff is the latter. Almost all of the strategies are price-driven rather than fundamental modelling of the sport. Like the Robot Wealth guys like to say: focus on the easy games. Most of the strategy ideas are things I’ve picked up over my career although one or two are thanks to Paul and two are strategies abandoned by my existing syndicates that I’ve managed to resurrect after getting the blessing of my colleagues and a lot of hard work.

In total 9 strategies made it to the point of being assigned a numeric code (i.e. interesting enough to be talked about publicly) but only four are actually still running. Interestingly, almost none of the initial strategies I had planned to work on going into 2025, the ones that gave me the confidence to do this, survived. I think in most cases they still have legs, they just need (a lot) more work to get them over the line. Strategy2, for instance, has been through four reimaginings and I have a very clear idea of what I want to do for the fifth but I’ll need to brush up on the Kalman Filter and, honestly, who has the time to do that?

Indeed, prioritisation was a key factor in making the year as much of a success as it turned out to be. Performance of the portfolio only really took off once I cut what wasn’t working and focused on what was. Stat arb style strategies were out, along with hard-to-monetise “writing lottery tickets” ones, and simple trigger-based strategies were in. Stuff like “lay a horse as soon as it falls”, where the “alpha” is brutally simple and what matters is the execution. These strategies don’t necessarily even have an idea of what the fair probability of a horse is, only that the current price is value.

Scaling up was another painful learning experience. It provided a one-two punch of teachings, firstly in that some of the more high variance strategies got off to a flying start, giving a misleadingly promising impression of their long-term performance. This was particularly the case for Strategy6 which, at one point during its over-exposed stage, had the highest total return of the portfolio.

The second lesson was that total returns didn’t scale linearly with stake size. It may seem incredibly naive to think that, and I certainly thought it would only be true up to a point, but that point proved to be much sooner than I anticipated. Aside from narrowing my focus to a core set of strategies, the other single biggest step forward came from reducing stake sizes and switching to more intelligent ways of putting more money through the market. Continuing with the example of laying horses as soon as they fall, that may equate to simultaneously backing some of the other runners, who will benefit from less competition, or also laying the horse in the place market.

We’re blurring the line now between improved execution and improved alpha but one could also try to pre-empt the signal, perhaps by identifying when a horse is approaching a jump too slowly, or at the wrong angle, or whatever other factors may influence jumping success. I don’t know, I’m not a horse racing expert, but we will be digging into this in a future article.

Looking Ahead

Overall, developing these new strategies turned out to be way harder than I expected going into 2025. It took me until November, possibly October, before I had a portfolio that could consistently make money week in, week out and, even then, the scale of the returns were far off what I was hoping. The headline result, that the new strategies made 7% of the total net return of the existing ones, overlooks the fact that this was a bad year for the existing strategies. It was arguably a good result but definitely for the wrong reasons.

I did many sports-betting-related things in 2025 which weren’t working on either my syndicates or personal trading. I pursued potential collaborations in South African horse racing, pre-event tennis, and trading on Polymarket. None of these have yet to bear fruit but they’re all things I’m hoping to progress in 2026.

I also ended up getting a part-time job during the summer. I’m now working for Amelco two days a week as a member of their quant team under the leadership of Milt Mavrakakis. Milt and I have been in each other’s orbit for many years now, typically meeting at Royal Statistical Society conferences, having deep respect for each other’s abilities, but never having had the chance to work together until now. The motivation for getting this job was three-fold:

1. Financial stability. Any professional gambler who’s tried to get a mortgage will understand the benefit of having a normal salary

2. Getting me out of the house. Being a professional gambler is an isolating experience and I’ve missed socialising with colleagues and feeling part of a team. I do work with other people in my syndicates but we’re dispersed all over the world, we naturally work remotely, and are generally siloed in our responsibilities

3. Exposure to new ideas. I think I’ve talked enough about the importance of this but over and above that, it’s one thing to try to make yourself try out a new technology or read an obscure research paper and another to naturally be exposed to these things through the course of your job. It’s also hard to conceive of how else I could have the opportunity to learn from a former employee of one of the most successful betting syndicates in the world.

In terms of those points, the job has delivered in spades and I’m very happy with my decision. It has, naturally, however, taken time away from working on the new strategies and I think it’s impossible to completely avoid the framing that lack of success with the new strategies led to me exploring other opportunities like this.

I’ve heard mixed things about what others made of 2025. For some, it was another data point in the inexorable decline of professional gambling: more competition, higher fixed costs, increased regulation and other factors leading to lower liquidity. For others, it was their best year yet. I’ve also met some new entrants who have really hit the ground running. Anecdotally, it seems like high frequency options trading really sets you up for success in professional gambling. As for me, I’m very much in the first camp. But I’m not out. Yet.

It’s impossible to prognosticate about professional gambling in 2026 without mentioning prediction markets. I’m increasingly of the opinion that this is where I should be spending most - if not all - of my time in the coming year. They have the feel of the early days of Betfair - no Premium Charge, often no bet delays, and easy opportunities everywhere you look. Access for UK residents is tricky but far from impossible, and while there’s the looming spectre of legality, I doubt this is existential. Nepotism is great, isn’t it? Regardless of the long-term outlook, I think it’s very much a case of making hay while the sun is shining.

The final thing to talk about is Substack. I’ve toyed with the idea of starting a Substack for probably two years now. My original plan was to use it as a jumping-off point for publishing a text book on, yes, algorithmic sports betting. I still aspire to publish such a book but the process of writing weekly updates about my new strategies in 2025, albeit on a different medium, has shown me that I have both an audience and something more to say.

Inspiration really struck during that festive lull at the end of 2025. I’d already decided that any future writings on my strategies would be done on Substack and that was fine for betcode members who already knew who I was and what I’m about but it felt like I needed to introduce myself to the rest of the world. I’d remembered that in the deepest darkest days of COVID I’d written up a bit about my background for just a handful of betcode members, the old guard if you will, craving the need for the human connection that was denied to us by lockdown. Disappointingly, I was unable to find a copy of that, no doubt a victim of Slack’s monetisation policies, but I am now certain it was for the best as it forced to recreate the piece from scratch and I’m delighted with the result.

So there will be Substacking in my 2026 future. I’ll still put out some free articles but paid ones will start making an appearance. Think deeply researched technical reports with genuinely actionable insights. Not pure alpha drops but things that wouldn’t be too hard to turn into actual strategies if you have the wherewithal. The price will be going up over time as the content library builds up so feel free to lock in a good price while you can. First up we’ll be talking about the distribution of goals in association football matches and why everyone gets it wrong.

You’ll no doubt note that I haven’t mentioned the new strategies in my 2026 plans. While I’ve talked about the importance of stripping back my focus to a core set of strategies that made money, this is something that can only be taken so far. It gets increasingly difficult to iterate, to keep having fresh ideas for how to push a strategy forward. It feels like I’ve taken these strategies as far as they can go for now. Certainly, the lower hanging fruit is elsewhere. I went into 2025 thinking my ideas list was like a to-do list: implement each strategy in turn and grow my bankroll exponentially. Not only was my sense of what it would take to get each idea over the line badly miscalibrated, but even my sense of what constituted a good idea was off. Coming to that realisation - and carrying that lesson into the 2026 battleground of prediction markets - is what I really earned in 2025.

Both of my parents worked in IT and growing up in the 80s we had a PC at home long before they became ubiquitous. One of my earliest memories is of my dad introducing me to a racing game on his 286 but I was much more interested in what happened when we quit the game: I was fascinated by what I would later learn was the DOS prompt and the latent power it suggested. This was a machine that could be controlled, commanded to do my bidding. Programmed.

I did well at school but there were limited scholastic outlets at that time for my ever-expanding interest in computing. I taught myself a range of programming languages from C++ to Visual Basic to x86 assembly. If I wasn’t programming, I was playing computer games. It seemed almost preordained that I would study Computer Science at university, all traced back to sitting on my dad’s knee. The only hiccup in this plan was a rejection from Keble College, Oxford. It must have stayed with me a long time, academic failure being an unfamiliar experience, because I only recently threw away the rejection letter almost twenty years later.

I ended up at Imperial College instead, in hindsight almost certainly a better option anyway. This tale is full of these “sliding doors” moments. I’m not particularly spiritual but I’m less inclined to think of these forks in the road as forever closing off opportunities and more that we ultimately end up where we’re supposed to be, simply having taken a more circuitous path to get there.

This first stint at university was challenging for me socially, flying in the face of everyone telling me to prepare for the best days of my life. Everything worked out OK in the end but I do wish I could go back and tell that to that younger version of myself. Academically, things were better but I still found it hard to fit in, just in a different way. The courses which most interested me were those on artificial intelligence and AI-adjacent like Statistics and Bioinformatics. I would take out obscure, Russian-authored books on neural networks from the now sadly defunct Mathematics department library, one of my former happy places but now just one of the many victims of skyrocketing South Kensington property prices. I would dream of constructing my own sentient machines, an almost comically transparent reflection of my own loneliness looking back on things. Writing this as 2025 comes to a close, I sound like a master prognosticator but this was the deepest depths of the AI winter and as far as my academic world-view was concerned, no one was studying or had any interest in neural networks. I was simply a decade too early.

My four year integrated Master’s course in Computer Science included a six month industrial placement between my third and final years of study. Most of my peers were headed to investment banks, a career path that clashed violently with the kind of left-leaning, idealistic and embarrassingly naive outlook on life I held at the time, before I knew any better. This is another of those moments where I look back and wonder how much further my career would have progressed if I’d gotten into the business of trading markets at such a young age. Instead, I decided that what I really wanted to do was develop computer games. Unfortunately, the UK game development scene was dead at the time; the first iPhone was about four years away from being released; and the concept of “apps” did not exist. This time I was not quite a decade too early but I might as well have been.

Instead I ended up at Philips, working on software for integrating GPS into mobile phones. My defining memory of it is working with a colleague who had reached that age where one either cannot hear one’s own farts or one simply stops giving a fuck and audibly lets rip whenever the need takes one. I’m looking forward to living such a carefree existence in the future.

While the industrial placement did little to set me up for a future career as a professional gambler, my Master’s thesis, conducted in my final year at university was my first introduction to the intersection of computers and betting. Entitled “No Limit Texas Hold’em Poker Robots” and supervised by the incredible William Knottenbelt, I was training neural networks using distributed computing to play poker almost twenty years before Pluribus. It may sound like this time, finally, I’d timed things right, but the hilarious thing is that Will went on to supervise students studying both betting exchanges, like Edmund Noon’s seminal work on extending the Kelly Criterion, and also predictive models for tennis, both way more relevant for my future career path than poker. So, once again, with Betfair still in its infancy, my time had yet to come.

I graduated with a first-class degree in Computer Science from Imperial College and a complete lack of direction. I’d made a half-hearted attempt to apply for a neuroscience doctoral programme at Edinburgh University that went nowhere, and no attempt to apply for anything resembling a real job. World of Warcraft was doing its best to derail my life, having been released in my final year at university. Thankfully, my dad got fed up with his feckless son and got me a job working as an IT technician in the school where he was teaching. I will be eternally grateful to him for showing me just how utterly, diabolically shit my life could be, and in short order I had applied for and been awarded a place on a prestigious Master’s plus PhD program in Bioinformatics at Imperial College, sponsored by the Wellcome Trust (the UK’s richest charity for those of you keeping score). That lacklustre application to the course in Edinburgh? I used that experience to absolutely nail this one. One door closes and another one opens.

I had a fantastic time at university second time around. Having been handed a do-over, I was determined to put myself out there, joining the fencing club and ultimately captaining the men’s second team for two years in a row. That taught me a lot about leadership. Not about being charismatic, or motivational, neither of which fit my unassuming personality, but about the practicalities of administering and running a team. In terms of the course itself, the idea of the programme was to take computer scientists such as myself along with mathematicians, physicists and so on, and teach us biology, while at the same time taking biologists and other life scientists and teaching them how to program, turning all of us into interdisciplinary scientists capable of advancing the study of biology using computing. That formed the Master’s part of the programme, then we were free to pursue a PhD in any of the relevant departments. I found the Master’s incredibly successful at one thing: finally beating the idealism out of me. I was no longer interested in dedicating myself to a life of scientific study for the betterment of humankind.

I found myself in a bit of a pickle. I was committed to doing a PhD, and a PhD in Bioinformatics no less, but with no interest in ultimately becoming a bioinformatician. It seems odd writing this now, but at no point had I considered simply dropping out. I was receiving an incredibly generous stipend (tax-free, a perk I guess I got used to), relatively speaking, and there were copious other funds available to pay for computer equipment and other research necessities. My idealism firmly shed, I was starting to think more pragmatically - some might say “mercenary” - and I figured that Statistics was a good subject area to focus on, being very widely applicable. I also had a sense of wanting to overcome a challenge. I think Statistics is a very unintuitive discipline, and I wanted to prove that I was capable of conquering it.

Ultimately I did conquer it, at least in the sense of being awarded a PhD in the subject. I learned a lot, really a lot, travelled to conferences in Brighton, New York, and Helsinki, and published a lot of papers. I also experienced what it’s like to be in an abusive professional relationship and, again, I wish I was able to go back and tell past Mo that everything was going to be OK. Doing a PhD can be a very isolating experience and I have a lot of respect for everyone that is able to make it to the finishing line.

It was on the way back from that conference in New York that I experienced a pivotal moment in my life, not that I knew it at the time. The plane had landed and we were all queuing in the aisles waiting for them to open the doors so we could finally disembark. I was enjoying being connected to the internet again and catching up on emails when I noticed a job advert that had been sent to the Statistics department mailing list. It was advertising a job building statistical models to predict the outcomes of sports events. This really caught my interest. I was fully prepared this time around to go work for a hedge fund but my brother and I had recently discovered a website called “Betfair” that offered an API for betting on sports and my eyes were starting to open to more outside-the-box ways of making the kind of money I now craved.

I applied for the job and ended up getting it, bonding with my future boss in the interview over our love for fantasy novels. I remember I was reading Steven Erikson’s Malazan Book of the Fallen series at the time. My new boss had recently been forced out of a betting syndicate where he’d worked as the head trader and so was not a technical guy but he’d seen enough of how the sausage was made that he figured he could impart his knowledge to his newly hired quant (that’s me) and let them work out the details. That’s exactly what I did. There were two financial backers and we started with tennis.

Why tennis? There will be many, many articles on this in the future but in brief, we’re talking about betting while matches are taking place. Every point presents a new betting opportunity. Not to bet on the outcome of the point itself, but to update your prediction of who is going to win the match based on new information, and to bet accordingly. Tennis features a highly structured scoring format. At any moment in time you’re in one particular match state and there are only two other match states you can transition to depending on which player wins the next point. It takes over a hundred points to win a tennis match so there are no sudden shocks. You get a pretty slow and consistent flow of information as to which player is ascendant. Contrast this to association football where a referee could give a penalty which is not a penalty and completely change the course of a match (perhaps less likely these days with VaR). Team sports are also naturally more complicated than individual ones. This may be mind-blowing if you only remember tennis exists when it’s Wimbledon but there are tens of thousands of tennis matches a year to bet on. That turnover of events really helps to maximise profit.

Development of the tennis model went well, leaning heavily on Tristan “The Baron” Barnett’s PhD thesis. We got one of my friends to do some contract work writing our trading software. I worked out that the official data feed, in those days provided by Enetpulse, encrypted their data using a Caesar cipher that was ultimately easy to crack. And then we started placing bets.

At some point it had been communicated to me that I would get a share of the betting profits doing this job. I honestly can’t remember if that was in the job advert or in the offer letter or what. From what I can recall, the key thing is that the amount, the percentage, was not stated up front. So we’d reached this milestone where the bets were being placed and the profit share offer was presented and it was low. Like single digit percentage low. Remember, there are four of us involved at this point plus my friend who had only done contract work without any expectation of a profit share and too busy doing his own projects - also in sports betting - to care. So there’s myself, my technically illiterate boss, and two largely silent financial backers. I put it to my boss that I am the one person in this whole venture who could go away and do it entirely by himself. Remember, I’ve got the PhD in Statistics for building the models but I’ve also got an undergraduate degree in Computer Science and I’ve been programming since I was about 12 years old so while my friend did help us hit the ground running, I’d have been perfectly capable of also writing the software myself. All credit to my boss, and to the financial backers who were all good guys. They saw the reality of the situation and they agreed that we would split the betting profits equally between the four of us, after the financial backers had gotten paid back their initial investment.

The betting went well. It might be hard to conceive of if you’ve never developed a profitable betting strategy yourself but the turnover of events is so frequent and the capital requirements so modest that you can start with something like a £100 deposit and end up making millions of pounds without ever making another deposit ever again. That £100 snowballs until it reaches some kind of terminal bankroll, the market capacity if you will, and then everything you make on top of that is just money in your pocket. That’s basically how it went. We made money one week, we bet more the next. We made more the next week, we bet more the next. And so on.

I’d started this job in late 2010. Betfair had existed for about a decade by this point. If you looked at how hard it was to make money then compared to the early 2000s, you’d probably have thought the game was up. Premium Charge had been introduced in 2008, essentially a 20% commission rate for successful customers. They’d also added bet delays and cross matching, denying us access to juicy returns enjoyed by early adopters. But we started just before increased rates of Premium Charge were introduced in 2011, with commissions as high as 60% for the most successful customers. If you were unlucky enough to fall into that bracket I’m sure you would have longed for the days of 2010. The story goes on. Betfair introduced the sportsbook in 2013, signalling a shift in focus away from the exchange and the start of an ongoing decline in exchange volumes. If you’d seen that in 2013, you’d probably have thought that now, finally, the golden goose was dead. Spoiler alert: I’m still successfully betting on Betfair in 2025 and I can personally introduce you to many others doing the same.

With tennis being “cracked” it was time to look at other sports. Association football was the obvious choice, it being one of the deepest betting markets in the world. I decided to approach it the same way I approached tennis: review the academic literature, take the state of the art and improve on it where possible. At the time, state of the art for in-play football was Dixon and Robinson’s paper modelling goals in football as Poisson processes. It probably still is, given the drop-off in publicly available material as the true value of it has come to be recognised. The model is computationally expensive so we turned to the best hardware that Amazon Web Services offered. That proved to be our undoing.

Whereas everything had seemed so effortless with tennis, we struggled to make money on football. In the meantime, we were paying a lot for the compute. This is hardly a phenomenon unique to betting on sports but having high fixed costs with high variance income is a painful combination. We were ploughing our tennis profits into the football with nothing to show for it. It was a dispiriting time.

Eventually, we decided to give up. Not just on football but the whole enterprise. The tennis was still making decent money but it seemed like more of a side project that could be maintained while we all moved on to better things. It seems strange writing these words now. Why didn’t we regroup, double down on the tennis? I suppose we all had our reasons. It’s perhaps hard to articulate how scarring the experience of trying and failing to get the football working was. Managing emotions is an important part of all types of trading and while we may aspire to develop fully automated systems that allow us to completely abdicate responsibility for the trading decisions themselves, it’s hard, perhaps impossible, to fully detach from the results.

For the financial backers, they had their other business ventures to keep them busy. For my boss, I think he was simply bored. He wasn’t someone especially motivated by money, and I don’t think he loved the game of sports betting the way I had come to. So we went our separate ways.

I say I’d come to love the game of sports betting but at the time I wasn’t sure. Perhaps it was all just numbers and it didn’t matter where the numbers came from and I’d be able to find joy in any kind of data analysis? I set out to test this theory by getting a job in a different industry. That would also let me continue to run the tennis strategy on the side without a thorny conflict of interest. So I updated my CV. I printed business cards. I went to recruitment fairs. And I ended up at Wonga.

Yes, I’d gone from sports betting to sanctioned loan sharking. If you know Wonga at all, you probably remember them as a predatory payday loan company that charged exorbitant rates of interest. You’d not be wrong but none of that really occurred to me at the time I worked there. I viewed the company as a thrilling trailblazer in the nascent fintech industry and was excited to join the Decision Science team that was dripping with talent. Many of these former colleagues went on to do great things including founding unicorns.

Working at Wonga taught me some valuable lessons about the nature of business and how it contrasts with trading. My team was involved in building “scorecards”, credit risk speak for models that predict whether a customer will pay back their loan or not. We all had a hard on for machine learning but this was before deep learning was a thing so random forests were the model du jour. So we built this shiny random forest model that was better than the existing logistic regression one in just about every single way and it never went into production. Why? Two reasons. Firstly, we had to worry about the regulator. We had to be prepared to explain why a customer was denied access to credit, or why we gave a loan to a customer that should never have been given one. That’s much easier with a linear combination of factors than a forest of hundreds of decision trees - and this was several years before SHAP. Secondly, responsibility for implementation lay with a separate development team. No doubt this is a common setup in many companies, those involved in trading included, but I’ve never seen it work well as there’s too much lost in translation.

Another thing that was completely baffling to me was we had access to this dataset provided by the credit bureau that contained information on things like default rate by postcode. Of course, this feature was highly highly predictive. Just don’t give a payday loan to people living in the most impoverished area of the UK, right? Wrong! Here comes the regulator again, swooping in to ensure “fairness”. It’s interesting to reflect on how such wilful ignorance of science is impacting societal decay across the Western world today but I’ll save that discussion for another article.

All of this made me realise that I missed the purity of sports betting. There’s no office politics, there’s no regulator breathing down your neck, there’s no separate development team busy circle jerking and calling it Scrum. There’s only one KPI that matters: profit. So, after less than 9 months at Wonga, I started looking for a new job, back in the sports betting industry.

I still had my little conflict of interest to worry about. The tennis was still ticking along nicely. I worked with an outstanding recruiter who I subsequently used when I needed to do some hiring myself.

One place I interviewed at was Spreadex, a company that strangely offers both sports and financial spread betting. They gave me a technical test which happened to quiz me on Markov chains in tennis. I obviously smashed that and, unsure what to make of me, they had the CEO make an unscheduled drop-in to my interview. I decided to pass for a couple of reasons. Firstly, I didn’t feel like there was actually a suitable role for me at the company. Secondly, I wasn’t sure that I wanted to switch sides in the sense of going to work for a bookmaker instead of prop trading. I’d have all of the above crap to deal with, just at a job that happened to be in the gambling industry.

I also interviewed at Smart Odds. If you’re not aware, they’re one of the biggest gambling syndicates in the UK, owned by Matthew Benham who owns Brentford football club. This was much more the kind of company I was looking to join but the impression I got was that you’d end up being “far away” from the money. What I mean by this is that you’re quite detached from the actual trading. It’d be much more of a pure academic research role, the output of which may or may not be acted upon in ways that you are not privy to.

I ended up joining SportsRISQ, a fairly unique company that specialised in prize indemnity insurance in sports but also did a small amount of prop trading that they were looking to expand on. I was brought in to build out a quant team which was an exciting opportunity to run my own team for the first time. They were also happy for me to keep doing the tennis on the side.

At this point in my story I’m going to have to start glossing over some details. We’re getting closer to the present day and there’s some secrets that aren’t mine to share.

I suppose like every role in my career, working at RISQ was a formative experience. In one, somewhat superficial way, I entered a Java developer and left a Python one. The CTO was a very interesting guy, former banker like most of the staff, and as well as Python he introduced me to Erlang and, in a more abstract sense, actor systems and, for want of a better word, micro-services. This was all highly influential in trading systems that I would go on to build.

I ended up recruiting four people to my quant team, most of whom I still work with to this day. One of them was my brother, who was our “in-house” developer. People often ask me what it’s like working with my brother, with a deeply sceptical twinkle in their eye, as if they know with absolute certainty that it’s a shit show which will end in tears, and I think to myself that they should really not project their own familial emotional baggage on to other people. So I smile and tell them it’s great, which it is. I’m guaranteed to speak to my brother at least once a week which is handy now he lives in South Korea, and he’s a developer and I’m a quant so we can largely each stay in our lanes without heated arguments about the best way to do things, and trust each other to get on with things and do a good job.

During my time at RISQ, the tennis stopped making money. The limitations of our trading software were starting to show. One example of this is that our flow was as follows:

1. Receive notification that a player has won the point

2. Poll Betfair for the latest market prices

3. Decide whether to bet given the updated model prediction and the latest prices

That poll was killing us. We had no way of asking for a price that didn’t already exist in our latest market snapshot. We had no way of asking for a price that may or may not exist in the market and simply seeing if we got filled. It seemed so naive in hindsight but we had designed our software for a world that no longer existed. Milliseconds now mattered and that poll was costing us milliseconds with no way of removing it.

You’re probably incredulous. Isn’t it just a simple refactor? You’ll have to take my word for it that it wasn’t. That flow was utterly fundamental to the entire trading software. My friend who’d written it had moved on to bigger and better things. (What about streaming? Still a year or two away). I guess I’d done the same. Things were going well at RISQ and I felt like that was where my future earning potential lay. I told the guys I was out.

It’s a story as old as time. We’ve heard a version of it already in this telling. Team does all the work to build a product. Management fails to adequately compensate them. Team leaves to make the product themselves.

That’s how my time at RISQ came to an end. From my perspective, certain promises were made that were not met. If you asked the other side, they’d deny those promises were made in the first place. It doesn’t matter now. As my former boss said: at the end of the day we made each other a lot of money.

I left with my brother and another team mate in tow. We hooked up with my friend who’d helped out with the tennis software way back at the start of my sports betting career. I still work with them to this day. That’s all I’ll say on the subject for now.

Let’s fast forward to the start of 2025 and get this wrapped up. Up to this point I’ve been working with various betting syndicates and I’ve done some contracting for other betting companies. I now start doing something I’ve never really done before: I start trading my own betting strategies.

Wasn’t the tennis my own betting strategy? In some sense, yes, but my friend wrote the software, and I split the profits with three other people, each of whom were involved in the day to day running to a greater or lesser extent. I’m now talking about being wholly responsible for developing the strategy from conception to implementation, risking my own money, not sharing the profits with anyone else, and not having anyone else to lean on for emotional support.

If you’re wondering what took me so long, it’s because I’m incredibly risk-averse. That might sound strange for a professional gambler to say, and one who left RISQ in the way that I did, but it’s true. I despise gambling when I don’t have an edge, and I want that edge to be large and to be as close to certain as it can be.

I also think I’m very bad at coming up with ideas. I lack imagination. But towards the end of last year I had something of a revelation: ideas are everywhere. Ideas are easy to find. There are more ideas out there than I have time to implement and trade. The hard thing is not coming up with the ideas, but turning those ideas into working strategies. And that is something I am good at.