How can we improve the prediction of asset returns with Machine Learning ?
Always looking for improvements... Happy and impatient to get your ideas !
Here is one of mine :
One easy way to do is by customizing the LOSS FUNCTION of the machine learning algorithm...
All errors in prediction do not have negative results*: if the error triggers a wrong decision, there is a loss (or an opportunity cost). Otherwise, the error has no financial consequence...
Therefore, it is useful to translate this asymmetry into the algorithm itself !
Want to know more about the Why and How, don't hesitate to read https://ssrn.com/abstract=3973086
Waiting for your views...
* Let's compare 2 algorithms and 1 investment strategy:
Algo A predicts 2 * the actual return (2% if the actual result is 1% and -1% if the actual return is -0.5%)
Algo B predicts 0% all the way (like with vanishing gradients...)
Investment strategy : we invest if the predicted return is positive, we don't invest if it is 0 or negative.
Even if the error is significant, Algo will lead to the perfect investment strategy with no wrong investment AND no missed opportunity
Algo B has exactly the same error in prediction: (2x - x = x - 0), but never invest !!!
Even if the 2 algos make the same error in absolute value, one is a perfect algorithm, the other is just worthless