quantlab.model.predefined.gats

GATs: Qlib’s graph attention model over one bar’s cross-section.

GATsNet reproduces Qlib’s GATModel (qlib/contrib/model/pytorch_gats_ts.py): a recurrent encoder per symbol, one fully connected attention head over every symbol of the bar, a residual and a two-layer output head. GATsRegressor is the TorchModel head that trains it with Qlib’s Alpha158 benchmark settings. See docs/research/qlib-gats-master.md §1.

Classes

GATsNet	Qlib's GATModel with an output of any width.
GATsRegressor	Qlib's GATs on the cross-section of every bar.