Doing day-trading. Data consist of certain hours for each day. Let's say each day variable number of data points. Let's say we can open positions some time mid-day and then we close everything at the end of the day. One trade per instrument per day is allowed.
What would be the right Loss Function?
One approach is to do time slices, and do predictions for each slice, then divide by the sum of all predictions for that day. But that doesn't seem to work.
Basically, if we won't "normalise" it somehow then NN will just learn to bet infinitely large amounts all over the successful instruments. But instead we want it to learn to pick the RIGHT trade at the RIGHT time.
I feel I'm missing something very basic. Any help/tips appreciated! I'm using PyTorch.
Submitted July 15, 2020 at 04:50PM by avloss
via https://ift.tt/30doKtV
Stock Trading NTS 