Help Requested! PyTorch LSTM Model (Part 2)

As a side-question: how would an LSTM work on a non-continuous time-series? i.e. I only want to add a last-10 days sequence to the input if the latest closing price is 2 std away from the historic mean (breaches Bollinger Band). Or also if I wanted to add data from other stocks.

Nice progress - figured you would run into this issue soon as it's one of the nuances with financial data and what makes algorithmic trading challenging 😊

Couple of suggestions:

• Standardize your features to N(0, I). Most neural networks work best when the scale of each feature is the same
• Don't use Dropout directly with RNNs. To regularize RNN properly, there are some variants (i.e. Watt and Plessis (2019), Zaremba et al. (2015))
• Architecture (depth and width of the layers) is very important and requires careful tuning. I'd suggest getting rid of all regularization at first, and start with a sufficient model capacity (enough depth or width) to overfit on the training dataset so you don't get constant values. Then start adding in some regularization and decrease model capacity carefully by checking performance on validation set.
• Returns are generally stationary with long-run mean approx. at 0. It is possible that the best predictor given your features is simply a straight line through 0. Try creating other features from market data or incorporating alternative data.
• Your idea of working with irregularly sampled time series (instead of trying to train a model that works in every time period) is a good direction to explore. There are variants on RNNs that can handle this (i.e. Schirmer et al. (2022), or this dissertation survey Shukla (2021))
• To add data from other assets, simplest way is to just add them as columns to your X's, with a vector-valued output Y