ice-skaters (docs)

skaters on a river: calibrated forecast features for streaming machine learning, in Python and in the browser.

Live demo · The paper

The idea, from zero

Suppose data arrives one row at a time and a model must predict, then learn, then move to the next row: sensor readings, poll numbers, prices, counts. That is streaming machine learning, and river is the standard Python library for it. The usual hygiene is to standardize each number on the fly, subtracting a running mean and dividing by a running deviation. That helps with scale and nothing else: one fat-fingered reading still arrives at full force, and worse, it poisons the running mean and deviation that every later reading is judged by.

ice-skaters replaces that with a stronger contract. Every numeric stream gets its own tiny online forecaster (from skaters, a zero-dependency forecasting library), and the model is handed the forecaster's two-number summary instead of the raw value: 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, on a universal scale where 2 is notable, 4 is remarkable, and 7 cannot be exceeded 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 and use

pip install ice-skaters

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 that does the two-number substitution for the input streams. LaplaceTarget wraps any regressor, in the style of TargetStandardScaler, to add the target's own pair, which a transformer cannot do since it never sees the target; the target itself stays raw. Both estimators pipe, pickle and deep-copy like any river estimator. Non-numeric values pass through untouched, and NaN is imputed by the forecast itself with z = 0: the model receives "expected value, no news" instead of a poisoned pipeline.

If you adopt one thing, adopt the wrapper. In the ablation, LaplaceTarget alone beat river's recommended (lag-free) pipeline on three of four of river's own datasets untouched and 10/10 under simulated feature contamination. The fairness audit then asked the referee's question, what about a baseline with raw target lags, and the answer sharpened rather than broke the claim: hand-added lags are strong on smooth single-entity series, fragile everywhere else, and the wrapper on top of whatever features you have, lags included, improved the strongest configuration found on three of four datasets and sat within noise on the fourth. Add LaplaceFeatures when you distrust the features themselves.

What the evidence says

On TrumpApproval with river's recommended pipeline (progressive validation MAE, burn-in 100, examples/trump_approval.py):

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 30/30 seeds under every contamination type tested. There is also a theorem: pathwise regret transfer with measured constants and no calibration assumption, proofs and a numerical bound check in papers/regret-transfer.md, and the full write-up in papers/ice-skaters-jss.pdf. Protocols, harnesses and the losing rows live in the timemachines repo, benchmarks/RESULTS.md section 6.

Cost, measured: about 390 microseconds per stream per sample, roughly 900x StandardScaler. Right for polls, sensors, market bars and anything at human timescales; wrong inside a hot path at hundreds of thousands of ticks per second. And the cheap-rollup control keeps the pitch honest: a free EWMA version of the same (mean, surprise) pair captures much of the robustness, ties or beats the Laplace pairs as features on the smoothest real datasets, and loses by 14-26% wherever the stream has structure an EWMA cannot track. If cost matters, start there; the calibrated forecaster is the upgrade path, and on smooth series its standalone forecast remains stronger than any of these regressions.

Boundaries, stated plainly

• Distance-based learners (KNN) do not benefit: neighbour averaging is already spike-robust and the extra dimensions degrade the metric.
• Entity-interleaved streams (many units multiplexed into one key) want per-entity forecasters; a single forecaster per key is handicapped.
• If your heavy tails are signal rather than noise, taming them costs accuracy. Whether the extremes are informative decides the coordinates.

JavaScript

The same construction runs in the browser with no build step and no dependencies: docs/js/ice-skaters/index.mjs ports the river pieces the study used (StandardScaler, TargetStandardScaler, LinearRegression and the Pipeline learn-then-transform semantics, line for line, parity-tested against Python to 1e-9) on top of the vendored skaters JavaScript twin. The live demo is a stream, two models and a fat-finger button, all client-side.

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.