rolling_ols(target, factor, window, name=None, metadata=None)

Apply rolling_ols to long-form panel inputs.

Parameters

target : Panel | Graph
target argument. factor : Panel | Graph
factor argument. window : int
Positive trailing-window length in rows. name : str | None, default None
Optional graph-node name. A generated name is used when omitted. metadata : Mapping[str, Any] | None, default None
Optional metadata stored on the graph node.

Returns

Graph
Lazy single-output graph. Call .compute() to materialize a Panel.

Examples

import polars as pl

from bagelquant_core import Domain, Panel
from bagelquant_core.composer import rolling_ols

domain = Domain(calendar=["2024-01-02", "2024-01-03", "2024-01-04"], universe=["a", "b"])
left = Panel.from_domain(
    pl.DataFrame({
        "time": ["2024-01-02", "2024-01-03", "2024-01-04"] * 2,
        "asset_id": ["a"] * 3 + ["b"] * 3,
        "value": [1.0, 2.0, 4.0, 2.0, 3.0, 8.0],
    }),
    domain,
)
right = Panel.from_domain(
    pl.DataFrame({
        "time": ["2024-01-02", "2024-01-03", "2024-01-04"] * 2,
        "asset_id": ["a"] * 3 + ["b"] * 3,
        "value": [1.0, 1.0, 2.0, 1.0, 2.0, 4.0],
    }),
    domain,
)

result = rolling_ols(left, right, window=2).compute().data
print(result)

Notes

Inputs are aligned by (time, asset_id) before the operation runs.

Rolling calculations run independently for each asset_id ordered by time.

The model is fit on prior rows only and predicts the current row.