Neural networks

Back to Models and Features

Neural networks learn flexible nonlinear maps, including recurrent forms for sequence structure in longer panels.

Pass any model string below as Arm(model=...). Extra names an optional dependency, Scaling flags whether predictors should be standardized, and Tunable counts the hyperparameters the search space exposes.

Model string	Description	Input	Extra	Scaling	Recommended preprocessing	Tunable
density_hnn	Paper-faithful Density Hemisphere neural network with prior-DNN OOB volatility emphasis and OOB volatility rescaling.	supervised	deep	no	feature lags/trends are built before fitting; X and y are standardized inside each fit	4
gru	Torch-backed GRU regressor.	supervised	deep	no	handled internally: X and y are standardized inside each fit	3
hemisphere_nn	Bagged Hemisphere neural network with mean and variance heads.	supervised	deep	no	handled internally: X is standardized inside each fit	3
lstm	Torch-backed LSTM regressor.	supervised	deep	no	handled internally: X and y are standardized inside each fit	3
nn	Torch-backed feed-forward multilayer perceptron regressor.	supervised	deep	no	handled internally: X and y are standardized inside each fit	4
transformer	Torch-backed Transformer encoder regressor.	supervised	deep	no	handled internally: X and y are standardized inside each fit	3

Reference

• Models reference page for ModelSpec, ModelFit, and fit conventions.