Saltar al contenido principal

PreprocessingSpec

A PreprocessingSpec is a tiny dataclass that travels with every trained model artifact. It remembers which processors were fit and in what order so inference code can replay the exact preprocessing chain on new data without ever reaching back into the training-time configuration.

Why​

Qlib's DataHandlerLP applies an ordered chain of Processor steps (rank-norm, z-score, min-max, outlier clipping, etc.) during training. At inference time the same chain must be re-applied — otherwise the model is scored on data with a different distribution than it was trained on, which silently degrades live performance.

Until now this was only tracked implicitly (the handler config was expected to be re-instantiated exactly). PreprocessingSpec makes it explicit: the spec is serialised into the model pickle and reloaded when the model is served, backtested, or paper-traded.

Shape​

@dataclass
class PreprocessingSpec:
    processors_pickle: bytes                # fit state for exact replay
    processor_specs: list[dict]             # {class, module_path, kwargs}
    feature_columns: list[str]
    label_column: str | None
    handler_cfg: dict | None
    metadata: dict[str, Any]

Training-side usage​

from alphaswarm.ml.processors import PreprocessingSpec
from alphaswarm.ml.handler import DataHandlerLP

handler = DataHandlerLP(
    instruments=[...],
    learn_processors=[...],
    infer_processors=[...],
)
handler.setup_data()

spec = PreprocessingSpec.from_processors(
    processors=handler.infer_processors,
    feature_columns=[...],
    label_column="label_5d",
    handler_cfg={"class": "DataHandlerLP", "module_path": "alphaswarm.ml.handler", "kwargs": {...}},
    metadata={"dataset_hash": "abc123", "fit_window": "2020-01-01..2023-12-31"},
)

model.fit(dataset).with_preprocessing(spec)
model.to_pickle("models/alpha_v1.pkl")

Inference-side usage​

from alphaswarm.ml.base import Model

model = Model.from_pickle("models/alpha_v1.pkl")
spec = model.preprocessing_spec
if spec is not None:
    df = spec.apply(new_bars)       # replay the chain, no re-fit
preds = model.predict(df)

Serving-side usage​

All three serving backends (MLflowServe, RayServe, TorchServe) know about preprocessing_spec. The TorchServe handler in alphaswarm/mlops/serving/torchserve.py calls spec.apply(df) before model.predict(df) when the attribute is present.