Time Series Decomposition & Forecast

Analyze evenly spaced observations with classical additive decomposition and additive Holt–Winters. Choose a synthetic example or paste your data, choose the period, horizon and hold-out, then run. Parameters are not fitted automatically.

Numeric mode accepts one value per line, with decimal point or comma and exponents. CSV mode accepts exactly two columns (label, value), comma/semicolon/tab separators and quoted cells; mark a header explicitly. Empty records are ignored; invalid data and excess rows are rejected together, never skipped. Labels are preserved in CSV, while plots use zero-based indices. Sampling intervals, order and missing observations are your responsibility; dates are not interpolated or checked. Maximum 2,400 observations, 262,144 input characters, 128 characters per label and absolute values ≤ 10¹⁰⁰.

Decomposition uses a centred moving average; even periods use a centred 2×m average. Seasonal phase means are centred to zero; residual = observation − trend − season. Components use separate vertical scales and the same time range. The edges lack a centred trend and residual. These descriptive components use all observations, including hold-out; they are not backtest predictions.

Holt–Winters initializes level from the first season’s mean, trend from the mean difference between the first two seasons divided by the period, and seasonal offsets from the first season. α and β update level and trend; γ here is γ* in sₜ = γ*(yₜ − ℓₜ) + (1 − γ*)sₜ₋ₘ, with all three in [0,1]. Since initialization uses two seasons, fitted values begin after that window; they still belong to an in-sample fit. Future forecasts refit on all observations. Additive seasonality assumes roughly constant seasonal amplitude.

Hold-out forecasts fit only the earlier training data. At least two complete training seasons must remain. MAE and RMSE score only the withheld observations. Seasonal-naïve MAE repeats the final training season as a baseline: a smaller MAE is better. Hold-out zero disables evaluation. Repeated manual tuning on this same hold-out can overfit it; this tool does not select parameters or guarantee future accuracy.

CSV exports full-precision observed values, labels, train/hold-out splits, fitted values, future forecasts, independent hold-out forecasts and baseline, plus decomposition. Missing initial fits, edge trends/residuals and future observations are blank. Editing clears stale results.

No multiplicative or damped model, period detection, ARIMA or prediction intervals. Holt–Winters equations · Classical decomposition · Forecast evaluation

Only input values are remembered in this browser. Calculated results and experiment histories are not saved.