jaycheng113/skill-retrieval-mcp/timeseries-forecasting

Forecasting a seasonal series

Split by time, never at random. A shuffled split lets the model see the future, and the resulting score is meaningless no matter how good it looks.

Decompose first. If trend and seasonality are additive, a linear model on lagged features is usually enough; if the seasonal amplitude grows with the level, take logs before fitting rather than reaching for a bigger model.

Evaluate with rolling-origin backtesting: fit on everything up to time t, predict the next horizon, advance t, repeat. A single held-out tail measures one draw from a distribution and will mislead you about variance.

Always report against a naive baseline — last value, or last season's value. A surprising number of elaborate models fail to beat it.