HumachLearnLearn · explore · apply
Start learning
Evaluation, Robustness & Advanced ML

Time-Series Playground

Build trend/seasonality/noise, preserve temporal order, create lag-aware forecasts and compare future errors.

20 · Metrics21 · Thresholds22 · Explain23 · Robustness24 · Shift25 · Time series
Lab concept guide

What to observe while you experiment

Time-series prediction must respect temporal order. Trend, seasonality, autocorrelation and lagged features create dependence, so random train/test shuffling can leak future structure and produce unrealistic estimates.

MechanismGenerate trend/seasonality/noise, create only past-dependent lag/rolling features, train on earlier times and evaluate on later times.
Failure modeRandomly shuffling time points or computing rolling/normalisation features using future observations.
VerificationInspect timestamps in every split and verify each feature at time t uses only information available at or before t; recompute one forecast error manually.
Experiment deliberately
Compare a chronological split with an invalid random split. Predict which score looks better and explain the information advantage causing it.
Time order is part of the data. Training always precedes testing. No random shuffle is used for the forecast split.
Ready.

Temporal split and forecast

Forecast error by horizon

Method comparison