Zeitreihen-Zerlegung & Prognose

A small forecasting workbench for evenly spaced series. Paste numbers (or load an example), set the seasonal period, press Run analysis.

Decomposition. Classical additive: a centred moving average of length equal to the period estimates the trend; seasonal effects are the average of (y − trend) in each season, centred to mean zero; the residual is what remains. Use it to see whether a yearly or weekly pattern is real.

Holt–Winters (additive). Level, trend and seasonal state updated with smoothing constants α, β, γ. The fitted line is the one-step prediction; the red segment is the h-step forecast from the end of the series. Defaults (0.3, 0.1, 0.2) are a reasonable starting point — raise α if the level jumps, raise γ if seasons drift.

Backtest. Hold-out keeps the last k points unseen, refits on the rest, and reports MAE and RMSE on those k forecasts. That is the honest check; in-sample fit always looks better.

Limits. Additive seasonality only (not multiplicative). No automatic period detection, no ARIMA, no prediction intervals. Cap is 2 400 points.