AD's daily selections

[AustinDillon75]

Stick with it AD. You are approaching the experiment in the right way IMO. Sometimes, things will be frustratingly unexplainable.

Thanks for your kind comment. It's pleasing when people who are so good at what they do themselves offer that sort of encouragement and it's typical of this forum that all posters support and help each other this way. I certainly don't want this forum to see my frustration boiling over.

I very much am sticking with it - I just want to do it justice and at present I need to spend the time ensuring I've got the models straight rather than just presenting the forum with slapdash results. At present I've not even looked at today's cards but I will do. Remembering the models do not have any regard for the daily odds so if we are choosing it at 8am the model is picking it at 5pm too unless there's a dramatic going change for it to account for.

 

[AustinDillon75]

What the models had been doing was to inject a total of 256 different features and metrics into the selections. Once the features go in LightGBM then runs it all, but what happens is there are simply too many features for it to comprehend. So to try and get rid of all the noise, and possibly, to try and get round any suggestion of overfitting, my idea was is to prune the 256 features by taking each code daily and getting the model to focus on the most relevant 100 odd features, before re-running the model focussed on the most relevant metrics. Again, the good old backtesting has added 100 or so points to April and May alone.

So today, as predictable as the rain, today threw up 40 races and 21 wins and a total profit on the day of 29.40 units. At Huntingdon, we got the first five home and were denied by On Lovers Walk in the last, which was a BSP of 16.5 anyway and would have been quite the upset. I can't say enough that it doesn't rely on big priced winners and Carmels Phoenix at 6.6 BSP in the 19:42 at Cork was our biggest priced winner today.

Hope rekindled. By way of an example, the model today regarded the attached 100 features as those most important but it reassesses them continually so we do not go stale. The overall profits of every single bet now (over 900) is back to a ROI of 7.3% but the last two days even with the huge loss on Monday is working to an ROI of 12.9%.