Free Analysis Reports - feedback wanted

[Gcampb]

Sounds like you have a similar sort of approach to me, except I chose the attributes. I use past form against todays race setup with a few other pointers to assign probabilities, and then price these up to a 95% book. I focus on the top race meetings only though as these are where form stands up best, which obviously has its limitations and requires higher staking. I have found without incorporating market data its best to get on as early as possible, which again has its limitations. Night before often not an option but morning of either on exchange or bookmakers (in shop) if liquidity isnt there. If you get closer to the race, without factoring market in, you often find market moves have taken you away from one's you have rated highest and would have previously backed. Obviously could go both ways of course but form tends to be used by tipsters and punters so often are subjected to moves.

Have you looked at Betfair historical data? The free data gives you 1 minute tick data, so it’s possible to check for compression and best entry points. I found it was mixed when I tested though there are an endless number of simulations to run.

[StellaBot]

FWIW
I hope you do well
I think you could use your time
more productively and quicker without
going back 10 years data

[Gcampb]

Today’s Hoofs analysis report is up - https://admindone.com/30-01-26

Quick thanks to the commenter from yesterday on Reddit, who implemented a simple dutching strategy using the report and came back with a positive ROI, always good to see people testing ideas properly rather than forcing bets.

As always:

this is analysis, not tips
execution and discipline matter more than the model
not every race is a bet

dutch bankroll curve .jpg

[sniffer66]

FWIW
I hope you do well
I think you could use your time
more productively and quicker without
going back 10 years data

I think the comment on 10 year old data is fair if you are looking at an individual horse in isolation. However, with an ML model you are looking to model the relationships between various metrics (features) on a runner type and compare them to the rest of the field, in the hope of being able to predict an outcome with x% confidence

That's with the assumption that horse racing hasn't changed fundamentally in the last 10 years

Just a quick simple example of the top of my head. Say you're looking to model an in-play L2B strategy, where you are looking to beat BSP on the lay and then catch a drift to get your back matched. So you might look at features per horse around early\late pace, track, going\weather history, jockey\trainer recent results etc etc.
Then split your data carefully so you have no data leaks, where the model can know the race result in advance, and also split so you have separate training, test and validation data. You can also do some clever stuff using folds to prevent overfitting

Point the model at your training split 10 years, of data and train over many epochs to see if the model can predict, with any confidence, whether a runner type, compared to the other runner types in the race (you cant look at a runner in isolation, it should to be compared to the rest of the field) is likely to drop x% from BSP post off then drift back to y% higher. Its the repeat patterns you're looking for the model to find

Because you've fed so many races into the model the hope is that it can learn the patterns\relationships between the features to be able to predict, on previously unseen events, whether a horse is a likely L2B candidate. Patterns that a human would never be able to see in years.

Once you have the model at a decent level of accuracy you can take it and, run it against your unseen validation data and see if the model predictions hold up. Or normally don't, in my case

Then it's back to the drawing board, rinse and repeat

[Dublin_Flyer]

That's with the assumption that horse racing hasn't changed fundamentally in the last 10 years

I was thinking of any biggish changes that come to mind that may skew data or be causing the noise. The standardising of the stall draw whereby 1 is always on the inside is not applicable as it was way back in 2011. Lingfield and Southwell both changed AW surface from polytrack to Tapeta and Fibresand to Tapeta in 2018 and 2016 respectively which would be inside the ML times.

Otherwise wind surgery/wind operations only came onto race cards in 2018. Horses first run since being gelded can occasionally be missed in race card notes. Also a few NH tracks can be affected by low sun in winter and have multiple hurdles/fences bypassed on the day which is completely weather dependent on a clear sky which would be hard to account for without checking individual race reports.

[sniffer66]

That's with the assumption that horse racing hasn't changed fundamentally in the last 10 years

I was thinking of any biggish changes that come to mind that may skew data or be causing the noise. The standardising of the stall draw whereby 1 is always on the inside is not applicable as it was way back in 2011. Lingfield and Southwell both changed AW surface from polytrack to Tapeta and Fibresand to Tapeta in 2018 and 2016 respectively which would be inside the ML times.

Otherwise wind surgery/wind operations only came onto race cards in 2018. Horses first run since being gelded can occasionally be missed in race card notes. Also a few NH tracks can be affected by low sun in winter and have multiple hurdles/fences bypassed on the day which is completely weather dependent on a clear sky which would be hard to account for without checking individual race reports.

Then the question is how much will those changes impact results. Also, its how you effect your data split. Ideally you don't want to split on years as such i.e train on 2015-2020 and test\validate on the rest. If you can create a random 80\20 split using all years (as long as you prevent future leaks) then train on that random set I'd assume that would smooth out the impact of the changes.

The obvious problem with limiting your data set to everything post major change is that you reduce the size of the data you can train the model on.