quantlab.model.predefined.realmlp#
RealMLP regression head backed by pytabkit.
RealMLPRegressor is a TabkitRegressor that wraps
pytabkit.RealMLP_TD_Regressor, the RealMLP network with the tuned
defaults from Holzmüller et al., “Better by Default” (NeurIPS 2024). It
trains on the flattened (num_times * num_symbols, num_features) rows of
the factor panel and predicts future returns as [num_times, num_symbols,
num_labels]. A panel is an xarray.Dataset indexed by timestamp
and symbol. Early stopping is pytabkit’s own, driven by the validation
segment the pipeline holds out, and the stopping epoch is written to the
tracking run after training.
pytabkit is itself a torch library. On macOS, torch and xgboost bundle
clashing OpenMP runtimes, so a process that mixes them must set
OMP_NUM_THREADS=1 before importing either; the same applies here.
Classes
Predict future returns with a pytabkit RealMLP (tuned defaults). |