ice-skaters

skaters on a river: calibrated forecast features for streaming ML pipelines.

The idea

Every numeric stream is replaced by two scalars from its own online Laplace forecaster: the predictive mean, which is what the forecaster expected this value to be, and the standardized surprise z, which is how unexpected the actual value was, bounded near |z| = 7 by construction. The mean carries the level. The z carries the news. A wild observation can move the pair only so far, and that bounded influence is where the robustness comes from.

Install

pip install ice-skaters

Quickstart

from river import datasets, linear_model, metrics, preprocessing
from ice_skaters import LaplaceFeatures, LaplaceTarget

model = LaplaceTarget(
    regressor=preprocessing.TargetStandardScaler(
        regressor=LaplaceFeatures()
        | preprocessing.StandardScaler()
        | linear_model.LinearRegression()))

mae = metrics.MAE()
for x, y in datasets.TrumpApproval():
    pred = model.predict_one(x)
    mae.update(y, pred if pred is not None else 0.0)
    model.learn_one(x, y)

LaplaceFeatures is a river transformer for the input streams. LaplaceTarget wraps any regressor to add the target's own pair, which a transformer cannot do since it never sees y. Both pipe, pickle and deep-copy like any river estimator. Details in the guide.

What you get

On TrumpApproval with river's recommended pipeline, progressive validation MAE with a burn-in of 100:

The front-end pays a small toll on clean data and holds its footing when the inputs misbehave. In controlled simulation the same substitution beats raw features, a running z-score, a median/MAD winsorizer and a Huberised loss 30 seeds out of 30 under every contamination type tested. The full protocols and numbers are on the papers page, including the cases where the front-end loses, because it does lose some.

Relation to the stack

skaters does one thing: fast univariate distributional forecasting, stdlib-only, in Python or the browser. timemachines builds anomaly detection on the same calibrated surprise streams. ice-skaters is the bridge from those streams to river's estimator protocol, and nothing more.