Hey gang,

Writing a follow up to my first cry for help where we figured out how to reshape the input tensors into the correct shapes and accept multiple features. Thank you Adam W !

I have made some more progress in the research environment. The model now accepts multiple features!

The inputs are the day-to-day pct_change of open, high, low, and close. The output is the day-to-day pct_change of close at N+1.

However, I am running into an issue where the model only predicts pct_change values near zero. What can I do to mitigate this issue and improve the model's predictive power?

Top graph is actual vs predicted close price. Bottom graph is actual vs predicted day-over-day pct_change.


Solutions I have attempted from scouring stackexchange:

  1. added dropout layer
  2. increased hidden_size (chose 128 arbitrarily. not sure what the rule of thumb is)
  3. reduced num_layer (chose 2 b/c 3+ produced completely flat predictions)
  4. changed loss function to nn.L1Loss (didn't seem to change the results much)


Ideas I haven't tried yet:

  1. standardize features instead of using MinMaxScaler
  2. “put an L1-Regularization term into your Dense Layers” (don't know what that means)