I’ve traded horse racing markets on Betfair for over ten years, and lately I’ve started thinking more seriously about where AI could eventually take this.
I know AI and machine learning have already been discussed here, but I mean something well beyond the tools we have today. Imagine genuinely capable AGI becoming available for perhaps £100–500 a month. It could analyse years of Betfair data, TPD and video, understand market behaviour, come up with its own ideas, test them, code the bots and keep adapting them with very little human input – all at least as well as a very good trader.
What happens when hundreds or thousands of people can point something like that at every Betfair market, including the small, low-liquidity ones? Would new opportunities keep appearing, or would the competition eventually squeeze most edges down to the point where an independent trader could no longer make a living, unless they had better data, faster execution or some kind of exclusive access?
I’m asking because I keep seeing more forecasts suggesting AGI could arrive before 2030. They could obviously be wrong, but it no longer feels like something from the distant future. I’d be interested to hear what people running serious automated or ML systems think about it.
What happens to Betfair trading if cheap AGI becomes available?
People tend to overcook both the impact of AI and the threat from it. There are clearly things that are going to change, and I am already using machine learning and other tools to extract more information from data. But it has not fundamentally changed what I do.
I also read a lot of academic papers to understand what people are thinking and to see whether there is anything genuinely useful in them. The problem is that many academic papers are simply wrong when they try to describe how people actually participate in markets, whether those are financial markets or betting markets. There is often a very large gap between analysing a market from the outside and actually operating successfully inside one.
That is an important distinction.
I have been active in markets for such a long time that the data being used to train models already contains my activity. If somebody builds a machine learning model from years of market behaviour, then in some small way it is learning from me and from thousands of other experienced participants who have already shaped those markets.
So when people say that AI is suddenly going to arrive and discover how markets work, there is a slightly circular element to that argument. Much of what it is learning from is the behaviour of people who already know how to operate in those markets.
That is why I think some of the discussion around AGI is overdone.
I have very little doubt that AI will be enormously transformative for the wider economy. But I suspect the effects at an individual level will be much more nuanced than the dramatic predictions suggest.
You might find a better way to design a product, package it, manufacture it, route it through a distribution network or manage the inventory behind it. Across an economy, improvements like that can create enormous gains.
But at the end of the process, somebody still has to make the product, put it in a box and ship it.
AI can radically improve the process. It does not necessarily remove the process.
I also read a lot of academic papers to understand what people are thinking and to see whether there is anything genuinely useful in them. The problem is that many academic papers are simply wrong when they try to describe how people actually participate in markets, whether those are financial markets or betting markets. There is often a very large gap between analysing a market from the outside and actually operating successfully inside one.
That is an important distinction.
I have been active in markets for such a long time that the data being used to train models already contains my activity. If somebody builds a machine learning model from years of market behaviour, then in some small way it is learning from me and from thousands of other experienced participants who have already shaped those markets.
So when people say that AI is suddenly going to arrive and discover how markets work, there is a slightly circular element to that argument. Much of what it is learning from is the behaviour of people who already know how to operate in those markets.
That is why I think some of the discussion around AGI is overdone.
I have very little doubt that AI will be enormously transformative for the wider economy. But I suspect the effects at an individual level will be much more nuanced than the dramatic predictions suggest.
You might find a better way to design a product, package it, manufacture it, route it through a distribution network or manage the inventory behind it. Across an economy, improvements like that can create enormous gains.
But at the end of the process, somebody still has to make the product, put it in a box and ship it.
AI can radically improve the process. It does not necessarily remove the process.
Also, I should add I don't think that AGI or True AGI is going to be particularly cheap, and that there will be a massive incentive for organisations to retain it themselves.
But the existing models are getting better and better all the time. I'm currently running two models, and each month or quarter I can seem to do more and more with them. My impression is that there's a skill in using them. I've worked with somebody who really struggled to get any utility out of them, but it's because they're not thinking about it in the right way.
But the existing models are getting better and better all the time. I'm currently running two models, and each month or quarter I can seem to do more and more with them. My impression is that there's a skill in using them. I've worked with somebody who really struggled to get any utility out of them, but it's because they're not thinking about it in the right way.
-
ShaunWhite
- Posts: 10878
- Joined: Sat Sep 03, 2016 3:42 am
AGI won't fundamentally give you predictive information that sophisticated ML can't already extract, although it may make doing it vastly easier. ML has been used in betting markets for years; I use TFT modelling myself.sudek wrote: ↑Tue Aug 25, 2026 4:58 pmI’ve traded horse racing markets on Betfair for over ten years, and lately I’ve started thinking more seriously about where AI could eventually take this.
I know AI and machine learning have already been discussed here, but I mean something well beyond the tools we have today. Imagine genuinely capable AGI becoming available for perhaps £100–500 a month. It could analyse years of Betfair data, TPD and video, understand market behaviour, come up with its own ideas, test them, code the bots and keep adapting them with very little human input – all at least as well as a very good trader.
What happens when hundreds or thousands of people can point something like that at every Betfair market, including the small, low-liquidity ones? Would new opportunities keep appearing, or would the competition eventually squeeze most edges down to the point where an independent trader could no longer make a living, unless they had better data, faster execution or some kind of exclusive access?
I’m asking because I keep seeing more forecasts suggesting AGI could arrive before 2030. They could obviously be wrong, but it no longer feels like something from the distant future. I’d be interested to hear what people running serious automated or ML systems think about it.
But it depends what you're trying to do. If you're trading price movement, that's always going to exist. Horses aren't machines and their precise odds are unknowable. Different models will arrive at different prices and, as they compete for what they see as value, that creates the very push and pull a trader can exploit. They're also news-driven markets. If a horse is refusing to load, you can see that just as well as an AI can.
The interesting implication is actually that more competing AGIs might make prices more efficient without making prices static. Those aren't the same thing and for a trader, that distinction matters.
I suspect the real edge gradually moves away from intelligence itself towards whatever remains scarce. Data, faster feeds and execution, lower costs, capital, queue position.
Being a few hundred milliseconds earlier may matter more than having a slightly better model.
Being a few hundred milliseconds earlier may matter more than having a slightly better model.
-
ShaunWhite
- Posts: 10878
- Joined: Sat Sep 03, 2016 3:42 am
10 or 20ms makes a difference already, and proven in simulations. It's why people use AWS Dublin with its 2ms ping and bespoke software to get an order out within microseconds of a price change. But the old chestnut is that it's not about being first, it's about being the first one who's right.
But there's all types of trader, auto, discretionary, fundamentals, ta, pure value, movement, pre play, in-running, different sports etc etc and plenty of methods so there's no one answer.
But people shouldn't get deluded about ML/AI, it doesn't find strategy for you, you still need a basic hypothesis to let it chew on and lots of data. And then you need a good working knowledge of the training parameters to use and various solution metrics, and a way to execute. There's no 'make me rich' button.
Once you’re competing for the same stale price or reacting to the same information infrastructure becomes part of the strategy but raw speed only creates value when the decision being accelerated has positive expectancy, being first to submit the wrong order just lets you lose faster. 
I think AI is overhyped, there are jobs it will never replace and certain human perceptual judgements that it can't replicate. But I do think it has somewhat a place on Betfair, like the next level of automation trading. Do automation traders who done this before AI feel threatened by it or that others can achieve what they have through an easier path?
On a side note I was recently pondering if Betfair trading as a whole is in decline. These days most of the younger generation don't use PCs or have the patience for such things. Would be interesting if PW could comment if there are less BA users nowadays. That would potentially be a net positive for traders unless the amount of genuine betting money is also in decline
On a side note I was recently pondering if Betfair trading as a whole is in decline. These days most of the younger generation don't use PCs or have the patience for such things. Would be interesting if PW could comment if there are less BA users nowadays. That would potentially be a net positive for traders unless the amount of genuine betting money is also in decline
-
ShaunWhite
- Posts: 10878
- Joined: Sat Sep 03, 2016 3:42 am
But it can find connections between things that humans can't. Such as processing thousands of maths papers and finding links between subjects where they might have used different perspectives or terminology. The 'sources that are wrong' is an issue but where models are being developed for a specific niche, firms spend billions on subject experts creating clean accurate training data. But yes the generic public models suffer from the quantity vs quality issue.
-
ShaunWhite
- Posts: 10878
- Joined: Sat Sep 03, 2016 3:42 am
No threat, I can't see AI being useful for live trading because it's too slow, and you'd need too many tokens to be feeding it updates every 20ms. The useful tool is ML, then you end up with a trained model that runs locally that can give you a judgement in milliseconds. AI might be useful for decisions that aren't time critical but otherwise what do you think it does that ML can't already do? What most don't have though is the data to train either AI or ML.
Maybe AGI is overhyped and will never match human intuition. None of us can know that for certain. I also get the feeling that some replies are looking at this mainly through what today’s AI and ML can do. I agree that current AI is too slow for a 20ms live loop and that ML still needs good data, ideas and an experienced user. But considering how quickly AI has improved, I’m not sure we can assume those limitations will always remain.
I also understand Peter’s analogy that AI may improve the design and distribution of a product, while somebody still has to manufacture it, pack it and send it. Trading would still require an account, capital, infrastructure and probably some human oversight. The difference, as I see it, is that the actual process from analysing information to placing a trade is digital and could potentially be automated from beginning to end, even if a person still owns and starts the system.
So putting AGI aside, consider a simple hypothetical question. What would happen if hundreds or thousands of traders as good as today’s best entered these markets and heavily outnumbered ordinary gamblers and less capable traders? Could the market become so efficient that there was almost no worthwhile profit left, even though prices continued to move?
Shaun’s point that the market will always move makes sense, but movement on its own doesn’t necessarily mean predictability. A series of coin tosses would also create a moving chart, yet the previous movement wouldn’t help us predict the next one. Horse racing is obviously far more complicated; I only mean that movement and predictable edge are not necessarily the same thing.
The stronger argument is that horses aren’t machines and there is no obvious fixed correct price. Different traders use different information and can arrive at very different estimates. But does that disagreement necessarily create patterns that someone can continue predicting profitably? Or could it simply produce constant price discovery, with prices moving as opinions change but almost no repeatable edge remaining after costs?
That is the hypothetical situation I’m trying to understand.
I also understand Peter’s analogy that AI may improve the design and distribution of a product, while somebody still has to manufacture it, pack it and send it. Trading would still require an account, capital, infrastructure and probably some human oversight. The difference, as I see it, is that the actual process from analysing information to placing a trade is digital and could potentially be automated from beginning to end, even if a person still owns and starts the system.
So putting AGI aside, consider a simple hypothetical question. What would happen if hundreds or thousands of traders as good as today’s best entered these markets and heavily outnumbered ordinary gamblers and less capable traders? Could the market become so efficient that there was almost no worthwhile profit left, even though prices continued to move?
Shaun’s point that the market will always move makes sense, but movement on its own doesn’t necessarily mean predictability. A series of coin tosses would also create a moving chart, yet the previous movement wouldn’t help us predict the next one. Horse racing is obviously far more complicated; I only mean that movement and predictable edge are not necessarily the same thing.
The stronger argument is that horses aren’t machines and there is no obvious fixed correct price. Different traders use different information and can arrive at very different estimates. But does that disagreement necessarily create patterns that someone can continue predicting profitably? Or could it simply produce constant price discovery, with prices moving as opinions change but almost no repeatable edge remaining after costs?
That is the hypothetical situation I’m trying to understand.
-
ShaunWhite
- Posts: 10878
- Joined: Sat Sep 03, 2016 3:42 am
It's a fair point that you also need a degree of predictability, and that's where pre--off sports events struggle. Financials have external forces that make them predictable, such as when yanks arrive in the middle east oil prices go up. Sports markets are so self contained and ethereal that there's almost nothing in the raw objective data to give strong evidence of the next move. Anyone who's into ML knows the real work is engineering new meaningful features from a very limited set of ingredients.
Ties in with what I've always thought, tech is fine and will make you money (for now) but if you want longevity then I'd swap all that for sharp subjectivity, that's where the tech has difficulties, processing the array of soft signals like rumours, the demeanour of a horse/player, vibes or human psychology.
Let's just hope that the people with the means to disrupt Betfair see their time as better spent disrupting something bigger.
Ties in with what I've always thought, tech is fine and will make you money (for now) but if you want longevity then I'd swap all that for sharp subjectivity, that's where the tech has difficulties, processing the array of soft signals like rumours, the demeanour of a horse/player, vibes or human psychology.
Let's just hope that the people with the means to disrupt Betfair see their time as better spent disrupting something bigger.
-
ShaunWhite
- Posts: 10878
- Joined: Sat Sep 03, 2016 3:42 am
I thought I'd give GPT this thread (and I might compare it with others like Deepseek later)....I haven't actually read it yet...
GPT :
I think the thread has slightly conflated two different questions: can AGI automate trading? and can enough intelligent agents eliminate trading edge? The second is the more interesting one.
If I were adding something genuinely new to the thread, rather than repeating what's already there, I'd probably post this:
I think there's another problem with the hypothetical: if you add 1,000 brilliant traders, you haven't just added 1,000 competitors, you've added 1,000 participants with different models, objectives, time horizons and positions.
Markets aren't a puzzle with a fixed amount of inefficiency which gets used up. The participants themselves create the market dynamics. An AGI trying to exploit another AGI changes the behaviour the other one is trying to predict, and both then adapt.
So I'd expect cheap AGI to compress simple, persistent edges very quickly, but not necessarily eliminate trading opportunity. It would probably change where the edge exists — away from obvious repeatable patterns and towards predicting the behaviour and interaction of the other agents.
Ironically, a Betfair market full of AGIs might become more complex rather than less.
And it continued...
That's the bit I think is missing from the discussion. Sudek's hypothetical implicitly treats efficiency as a destination: keep adding smarter participants and eventually you converge on it. But in a reflexive market, smarter participants alter the system they're modelling.
There's also an interesting consequence: if everyone gets essentially the same cheap AGI, diversity may actually decrease. Correlated models could see the same opportunity, act simultaneously, overshoot, withdraw liquidity simultaneously, etc. That can create entirely new exploitable structure rather than producing a perfectly efficient market.
So I wouldn't argue that AGI couldn't hurt existing edges — it almost certainly could. I'd question the assumption that lots of intelligence → equilibrium → nothing left to predict. Markets don't necessarily behave that way.
It's a good thought experiment because once you get past “AI will make everyone better”, the second-order effects — agents responding to agents that are responding to them — become much more interesting than the technology itself
GPT :
I think the thread has slightly conflated two different questions: can AGI automate trading? and can enough intelligent agents eliminate trading edge? The second is the more interesting one.
If I were adding something genuinely new to the thread, rather than repeating what's already there, I'd probably post this:
I think there's another problem with the hypothetical: if you add 1,000 brilliant traders, you haven't just added 1,000 competitors, you've added 1,000 participants with different models, objectives, time horizons and positions.
Markets aren't a puzzle with a fixed amount of inefficiency which gets used up. The participants themselves create the market dynamics. An AGI trying to exploit another AGI changes the behaviour the other one is trying to predict, and both then adapt.
So I'd expect cheap AGI to compress simple, persistent edges very quickly, but not necessarily eliminate trading opportunity. It would probably change where the edge exists — away from obvious repeatable patterns and towards predicting the behaviour and interaction of the other agents.
Ironically, a Betfair market full of AGIs might become more complex rather than less.
And it continued...
That's the bit I think is missing from the discussion. Sudek's hypothetical implicitly treats efficiency as a destination: keep adding smarter participants and eventually you converge on it. But in a reflexive market, smarter participants alter the system they're modelling.
There's also an interesting consequence: if everyone gets essentially the same cheap AGI, diversity may actually decrease. Correlated models could see the same opportunity, act simultaneously, overshoot, withdraw liquidity simultaneously, etc. That can create entirely new exploitable structure rather than producing a perfectly efficient market.
So I wouldn't argue that AGI couldn't hurt existing edges — it almost certainly could. I'd question the assumption that lots of intelligence → equilibrium → nothing left to predict. Markets don't necessarily behave that way.
It's a good thought experiment because once you get past “AI will make everyone better”, the second-order effects — agents responding to agents that are responding to them — become much more interesting than the technology itself