pretrain_pooled¶

Defined in fynance.models.objective

pretrain_pooled(model, Xs, ys, **fit_kw)[source]

Pretrain one ObjectiveModel on a pool of aligned assets.

The (X_i, y_i) pairs are pooled into a single training run so the net learns a signal shared across assets (transfer learning across a panel). The pooling is segment-safe: each asset’s series is kept as one contiguous chunk and mini-batches are drawn within a chunk, so no mini-batch ever crosses an asset join — the turnover carry (cost) and the temporal order that the single-asset fit relies on stay intact per asset. Concretely this extends fit’s contiguous chunking with segment boundaries: all segments’ chunks are pooled and (when shuffle) globally shuffled for SGD mixing, but every chunk stays inside its own segment.

Typical use: pretrain a shared net here, then adapt per asset with clone + finetune.

Parameters:

modelObjectiveModel: The model to train in place (its hyper-parameters and seed are used).
Xssequence of array-like: Per-asset feature matrices, each (T_i, F) (or a (T_i, N, M) panel); all must share the feature dimension.
yssequence of array-like: Per-asset realized returns aligned with Xs ((T_i,) or (T_i, N)); must have the same length as Xs.
**fit_kw: Per-run overrides of the training hyper-parameters epochs, lr, batch_size, shuffle, cost and seed (restored afterwards).

Returns:

ObjectiveModel: The same model, now pretrained on the pool.

Raises:

ValueError: If Xs and ys have different lengths, or the pool is empty.

Examples

>>> import numpy as np
>>> from fynance.models import ObjectiveModel, pretrain_pooled
>>> rng = np.random.default_rng(0)
>>> Xs = [rng.standard_normal((16, 2)).astype("float32") for _ in range(3)]
>>> ys = [(X[:, 0] * 0.01).astype("float32") for X in Xs]
>>> m = ObjectiveModel(layers=(4,), epochs=3, seed=0)
>>> _ = pretrain_pooled(m, Xs, ys)      # one net trained on all three series
>>> m.predict(Xs[0]).shape
(16, 1)