Value Betting model

[sniffer66]

There was a post recently about using ML modelling and selling the resultant model onto a syndicate. That went quiet but it got me interested in revisiting some of my old ML models and data. I've recently treated myself to an Nvidia 5060 ti GPU with 16gb VRAM and that can crunch through years of data in minutes using CUDA compared to my old CPU, taking days

I pulled out and updated my 12 years of pre off and post race data and trained a few models looking at value betting close to the off. I'm using a mixture of history data per runner and grabbing quite a bit more pre off price data until 10s from the start. That then gets passed to the model and model probability is compared to BF Implied probability. It then backs Top n runners based on my back testing (Max of 2 per race) where the model disagrees by x% - with a few filters for max\min price, runner count etc

Its not looking at individual horse data as such but modelling patterns, win rates, going , tracks and giving each runner a win probability based on the models learning. I'm using about 20 features\data points to achieve that.

I've done all the usual Train\Test data splits and tested again on Out Of Sample data so I know it's not backfitted

Put it live about 10 days ago and results are promising so far. For now its a £1 min stake live bet until I have some confidence. On Win , a 42% ROI is obviously ridiculously good so I expect a reversion back to around my back tested ROI soon. Place is ticking away, but given prices, it's not expected to be spectacular.

Drawdown on the backtest is acceptable so I'm happy to let it run in live for a while.

Will post updates as time goes on

output.png

Final value ≈ £93.90

ROI ≈ 42.3%

output (1).png

Final value ≈ £11.45

ROI ≈ 3.6%

[Euler]

The only thing I'd say to you, based upon my data and how my models are performing, is that favourites are winning slightly more often than they should be at the moment. You may want to take account of that if you're betting on favourites.

Or if you're betting on outsiders that may actually do the opposite and depress your returns, but it's worth making a note of this current situation.

[sniffer66]

The only thing I'd say to you, based upon my data and how my models are performing, is that favourites are winning slightly more often than they should be at the moment. You may want to take account of that if you're betting on favourites.

Or if you're betting on outsiders that may actually do the opposite and depress your returns, but it's worth making a note of this current situation.

Looking at the price spread its actually a mix. It's purely based on the model probability vs BF Implied and taking the largest gaps between the two, exploiting possible mis-pricing. Hopefully it keeps going as well as it's started

[Euler]

Fingers crossed!

[tico]

Hopefully it keeps going as well as it's started

It wont !! I almost guarantee it.Automation and value seeking don't work siimply because AI (and more primitive models that came earlier) cannot be subjective. It is the subjective element that dictates what value is .
There is a major movement towards automation ever since the first PC spun it's floppies and DOS was born.(to the detriment of us all)
None of it works .The problems has always been that the ability (and it is a profoundly impressive ability) for computers to gather and present data ,is somehow equated with it's ability to predict what really is a randome outcome .A computer will hazard a guess at value ,but it will be acting on a human definition ,and in my experience ,there are more human definitions of what value is in sport than there are stars in the Galaxy.It is possibly indefinable.
So the base variable has to be flawed .But it want stop glorious runs of winning ,but by the same token less-glorious runs of losing.
Computer technology (and this applies even more to AI) is a remarkable aid to winner finding ,but it wont find you any winners ,yopu have to find those yourself.
Regards
Tico

[sniffer66]

Hopefully it keeps going as well as it's started

It wont !! I almost guarantee it.Automation and value seeking don't work siimply because AI (and more primitive models that came earlier) cannot be subjective. It is the subjective element that dictates what value is .
There is a major movement towards automation ever since the first PC spun it's floppies and DOS was born.(to the detriment of us all)
None of it works .The problems has always been that the ability (and it is a profoundly impressive ability) for computers to gather and present data ,is somehow equated with it's ability to predict what really is a randome outcome .A computer will hazard a guess at value ,but it will be acting on a human definition ,and in my experience ,there are more human definitions of what value is in sport than there are stars in the Galaxy.It is possibly indefinable.
So the base variable has to be flawed .But it want stop glorious runs of winning ,but by the same token less-glorious runs of losing.
Computer technology (and this applies even more to AI) is a remarkable aid to winner finding ,but it wont find you any winners ,yopu have to find those yourself.
Regards
Tico

Yeah, I do appreciate that, but I like playing with this stuff, and have done for years. I ran a pretty successful Poker bot for about a decade back in the 2000's - but the human element in that made it a little different. I'm not expecting anything life changing to come out of it but it will be interesting to track how it does for pennies against the back test.

At the end of the day if its has more than a 2% edge in picking the right price versus the market then its a winner. Again, will be interesting to watch

On a side note, I started in IT in the days of Windows 3.1 and DOS 6.21, so get you on that. Having said that MS paid for my house, career and decent pension lol