quantlab.model.config.ModelConfig#

class quantlab.model.config.ModelConfig(factors, labels, model_save_dir, factor_data_strategy, label_data_strategy, start_date=None, end_date=None, hyperparameters=<factory>, val_size=0.2, metric_splits=('train', 'val', 'test'), random_seed=42, train_start=None, train_end=None, test_start=None, test_end=None, tracker=<factory>, name=None)[source]#

Bases: FrozenConfig

Config of every model head, torch or library.

It holds only what both variants read. Everything a variant or a head reads for training goes in hyperparameters, one flat dict. The base classes and the shipped heads read these reserved keys from it:

epochs: Epoch cap of a TorchModel; default 100, a positive integer.
lr: Learning rate of the default TorchModel._init_optim; default 1e-3.
early_stopping, early_stopping_patience: The library’s native early stopping in the shipped library heads; default off and 5 rounds (or the library’s own unit).
training_target: A LibraryModel’s training target: "cs_rank" or "cs_zscore", applied per bar; unset trains on the raw label. Set, label_scales is "standardized"; metrics still read the raw label.

batch_size, num_workers, pin_memory, prefetch, panel_device, panel_dtype

Reserved for the torch data loader and training panel.

A LibraryModel’s _init_model receives the dict without the library keys above. A TorchModel’s receives the whole dict, reserved keys included, so a torch head never splats it into a network; it reads its own keys by name, or drops the reserved ones with BaseModel.head_hyperparameters.

train_start, train_end, test_start and test_end bound the training and test windows; rolling cross-validation overwrites them fold by fold. start_date and end_date bound all the data the model collects; they are passed to every factor and label per request.

Examples

With factors and labels lists of factor objects built earlier:

>>> cfg = ModelConfig(
...     factors=factors,
...     labels=labels,
...     model_save_dir="/data/models/xgb",
...     factor_data_strategy="read",
...     label_data_strategy="read",
...     train_start="2018-01-01",
...     train_end="2022-12-31",
...     test_start="2023-01-01",
...     test_end="2023-12-31",
...     hyperparameters={"max_depth": 6, "early_stopping": True},
... )
>>> cfg.val_size, cfg.hyperparameters["early_stopping"]
(0.2, True)

end_date = None#: Last date of data to collect, inclusive. None means no upper bound.

factor_data_strategy#: "read" loads factor values from their stores; "cal" computes them first.

factors#: The factors whose values form the model’s input features.

hyperparameters#: Training and architecture settings, one flat dict; see the reserved keys above.

label_data_strategy#: "read" loads label values from their stores; "cal" computes them first.

labels#: The factors (labels) whose values form the prediction targets.

metric_splits = ('train', 'val', 'test')#: The segments a trained unit is scored on, any of "train", "val" and "test", in that order: their metrics and losses, and the bars its evaluation predicts. ("test",) scores only the test segment and predicts only its bars, which a walk-forward training repeats per fold.

model_save_dir#: Root directory checkpoints and their config.json are written under.

name = None#: Dotted import path of the model class; filled by the config setter.

random_seed = 42#: Seed applied to Python, numpy and, for torch heads, torch before training.

start_date = None#: First date of data to collect, inclusive. None means no lower bound.

test_end = None#: Last date of the test window, inclusive.

test_start = None#: First date of the test window, inclusive.

tracker#: Where training runs are tracked (ADR 0015); the default NullTracker sends nothing anywhere.

train_end = None#: Last date of the training window, inclusive.

train_start = None#: First date of the training window, inclusive.

val_size = 0.2#: Fraction of the training window held out, at its end, for validation.