1

What should be our target if we have our features and our return is variable. Suppose we have:

feature_1 feature_2 potential_payout actual_payout
       10        20              200          +200
        5         3               20           -10
        1         2                5           -10

So, we have this sort of a "gambling" set up. High potential payouts are being paid less frequently. How to model this with Neural Nets? I've tried many things like computing implied probabilities and modelling them directly, but that doesn't seem to work well (I understand there's a good reason for that). So what should be the target, does using NNs even make senes? Is there any already-made package for PyTorch? Please think of this like "which horse should I bet on" given past races. Is there any closely related Kaggle example?

avloss
  • 111
  • 3
  • 1
    Practically any regression model could be used here, and there is no reason to believe that NN would work better. My suggestion would be to try a couple of different models and see what is possible. – user2974951 Jan 10 '22 at 09:58
  • Thank you! I'm still struggling to see what should the Y be. Having "actual payout" as the Y - simply didn't work for me (perhaps I was doing it wrong). – avloss Jan 10 '22 at 10:05
  • 1
    @avloss it would be worth expanding on why you think having `actual payout` as the target didn't work. If you're not sure on 'why?' then try to include some information about 'how' it isn't working (plots of residuals for example). You should edit your original question to give us this info, more information means we can give you more specific advice! – jcken Jan 10 '22 at 10:08
  • I'm not entirely sure, it's just that validation was extremely poor. I've assumed I should do something like "log of implied probability" or "exponential" or some other "trick". But perhaps there's not "quick win" here. – avloss Jan 10 '22 at 10:20
  • 1
    Are you feeding `potential_payout` in as a predictor, or is this an intermediate output? I'm all with @user2974951, any simple regression model would be a reasonable benchmark. You may be interested in [How to know that your machine learning problem is hopeless?](https://stats.stackexchange.com/q/222179/1352) – Stephan Kolassa Jan 10 '22 at 11:02
  • No, it's a gambling problem. Like betting on a horse. You bet 10, you win 200 or lose -10. So it's just a question on which horses to be and how much. – avloss Jan 10 '22 at 11:13
  • Perhaps I've formulated it horribly wrong. Somehow it feels like this should be a pretty well explored problem. Yet somehow I can't formulate it properly. We have features of a horse, we have a payout (which supposed to be indicative of implied probability), and then we have outcomes (win/lose). Thank you for your help! – avloss Jan 10 '22 at 11:15

0 Answers0