Starting points for implementing ML/AI into sports trading

[eightbo]

Can't help but feel a tad behind when I see videos like the below.

https://x.com/ArchiveExplorer/status/20 ... 4791854194

What would you guys recommend as a starting point for implementing ML/AI into my trading?

[Euler]

I think a lot of posts like this aren't true or realistic. Of course the buzzword is AI at the moment and everybody's using it in every form possible but none of the strategies I've developed over the last 20 years have involved using machine learning or AI.

I'm not saying that you shouldn't give it a try or that you shouldn't explore new areas and so on. I'm just a bit sceptical, given the level of focus on these things, that people are using it the way they claim they are.

IMO If you learn to use some of the latest tools, you can use them to gather and analyze data much faster than you ever could have done in the past. I think that's a really unexplored area for a lot of people. That's where you'll find a few edges.

[jamesedwards]

You don't sell something you believe makes 200% per week.

Ask them to share a longer term P&L and see what happens.

[eightbo]

...
IMO If you learn to use some of the latest tools, you can use them to gather and analyze data much faster than you ever could have done in the past. I think that's a really unexplored area for a lot of people. That's where you'll find a few edges.

Thanks Peter, that's of great help.

I recall seeing you mention somewhere you almost doubled a year-to-year profit somewhere and was just so curious! I'll assume this sort of thing this was a contributor.

Thanks so much for starting BA. 22 years at the forefront of the exchange, and transformed so many lives with your team/product over the years. So grateful.

[eightbo]

You don't sell something you believe makes 200% per week.

Ask them to share a longer term P&L and see what happens.

It's more highlighting a proof of concept (code converting stream into TPD-like data)

I presume a few are on GPT-4o or Claude 3.5 Sonnet paid models? Let me know if you guys have gotten on well with anything else or that presumption is incorrect

[ShaunWhite]

It's more highlighting a proof of concept (code converting stream into TPD-like data)

I presume a few are on GPT-4o or Claude 3.5 Sonnet paid models? Let me know if you guys have gotten on well with anything else or that presumption is incorrect

Doing pretty well but haven't touched it for about 6 months cos I haven't been well/been busy. Back on it soon.

But two key things, #1 data obviously, as much as possible, and #2 having a clear idea about what you want it to do. It's not going to find a stragegy for you, but it can help you refine one. And execution is a factor too, ie are you going to go fully auto incl retraining, through to just daily notes, depends how much time you want to put into it. I was quite well placed with years of tick data and a sim/trading system so that informed my choices.

AI will help get the code together and streamline your process, there'll be a lot of iterations so it's best to be slick, and then choose an ML. I'm on TFT modelling (google it) but there's several, DeepAR, N-BEATS, TiDE etc etc.

One small glitch is that non of the temporal forecasting models understand that more than one selection constitutes a 'race' and the overround, but you can work around it.

Model training can be slow, anything from 15mins for a short data history with a handful of parameters to a several days for a few years of data, 100s of model params and 20 or more iterations (epochs).

Live price inference speed can be an issue too depending on the parameter library, from about 80ms for a simple model to 3 or more seconds for a complex one, so that affects how you use it. If you'll looking for best performance then you could easily spend several hundred a month on AWS GPUs.

It's good though, makes a change and takes away a lot of the tedious analysis.

[Kai]

I presume a few are on GPT-4o or Claude 3.5 Sonnet paid models? Let me know if you guys have gotten on well with anything else or that presumption is incorrect

My quant partner swears by the ChatGPT Pro and it keeps impressing, we've been using it to good effect recently

The freebie stuff would lose coherence instantly, it's only good for queries like "will ice cream give me brain freeze"

[sniffer66]

I presume a few are on GPT-4o or Claude 3.5 Sonnet paid models? Let me know if you guys have gotten on well with anything else or that presumption is incorrect

My quant partner swears by the ChatGPT Pro and it keeps impressing, we've been using it to good effect recently

The freebie stuff would lose coherence instantly, it's only good for queries like "will ice cream give me brain freeze"

I still have this argument with GPT 5.5 on a regular basis , then have to go back and correct it's code constantly !

Screenshot 2026-06-19 174526.jpg

[ShaunWhite]

I still have this argument with GPT 5.5 on a regular basis , then have to go back and correct it's code constantly !


Screenshot 2026-06-19 174526.jpg

It'll often revert to 'price' being a price like a share price, where higher is indeed worse. '(decimal) odds' leaves it in no doubt what you're on about. Using 'price' requires it to remember the context.

[sniffer66]

I still have this argument with GPT 5.5 on a regular basis , then have to go back and correct it's code constantly !


Screenshot 2026-06-19 174526.jpg

It'll often revert to 'price' being a price like a share price, where higher is indeed worse. '(decimal) odds' leaves it in no doubt what you're on about. Using 'price' requires it to remember the context.

I'll try that Shaun. I've told it to remember a good few times but it still loses the memory between chats

[Euler]

I presume a few are on GPT-4o or Claude 3.5 Sonnet paid models? Let me know if you guys have gotten on well with anything else or that presumption is incorrect

My quant partner swears by the ChatGPT Pro and it keeps impressing, we've been using it to good effect recently

The freebie stuff would lose coherence instantly, it's only good for queries like "will ice cream give me brain freeze"

I would often wheel in developers or PhDs to solve specific problems for me in markets that I was interested in exploring, but I no longer use either.

ChatGPT Pro on deep research is as good as any PhD in my experience. I can feed it an academic paper as well and it will make it coherent for me and allow me to understand and use some of the context in the paper.

On the coding front Claude Code or Codex can create the tools I require to analyse vast data sets really quickly and easily but also visualise and manipulate the data in a way that makes sense to me in a trading context.

This is how I'm using AI at the moment.

I now use transcription software which has become really good, especially in an environment where I've got live commentary and stuff going on in the background. It sped up my use of a PC dramatically.

[rosecruz]

I wouldn't trust that post, for two reasons.

First, Sportradar is a premium service aimed at bookmakers and other professional operators in the industry. Their
pricing isn't public — and when pricing isn't public, it's almost always because it isn't meant for the general
public. I found a similar service, https://opticodds.com/pricing, and it's the same story: you have to contact them
for a quote. Just look at the dropdown showing who they work with.

Second, that post was made on 3 March 2026, and ever since, that same account has done nothing but upload content
about AI models.

For context: I've been building ML models for horse racing for the past year — first spending months just gathering
data, and about three months ago finally producing a regression model that's genuinely quite good. Even so, if I
simply backed every top-rated selection it puts out, I wouldn't make a penny.

What the model has given me is a frame of reference for understanding the markets. Using it alongside watching the
market every day, I feel I'm finally starting to get the hang of certain situations. I'm building a sense of which
trainers the market favours and when, and the same with jockeys. For example, my model might rate a horse at a 2%
chance while it's trading at 2.0. These sort of discrepancies/agreements between model and market prices are the sorts of situations where I'm starting to find opportunities.

You seem to have more experience than me, so maybe the right approach is to start with Claude and build DuckDB
databases from your current data — they're lightweight and interact with Python really well. You don't need to know
any programming: Claude, Codex or ChatGPT will do all of it for you. If you can think it, and you've got the data to
analyse, your AI tool should help you write the scripts to analyse it — or even give you the answer by exploring the
data itself.

It's a huge amount of work, though, and honestly I don't think you can automate any of it from the outset. You have to
get your hands dirty with the data and be actively involved in the markets first. That's been my experience, at
least.

I personally love Claude. I've installed it locally on W11, and you can do amazing things with it once you've got the data. Just
remember to back everything up!

[tommyc333]

Another vote for Claude - apart from the usage limits which always seem to kick in when you’re deep into debugging something, it’s fantastic. I’ve used it for simple Python exercises, all the way up to building production websites.

Ultimately though, it’s not really the choice of tool which will make the difference, it’s how you utilise it.

DuckDb is great too, the only downside is that it only allows for one concurrent transaction and user / process. This means that if you are trying to import and export data at the same time it will lock and stop one of the processes.

For this reason I use PostgreSQL. It’s also free and interacts well with Python.