vault / models / forecasting.md

Forecasting

Modelling 60 words 1 outgoing 1 backlinks

Pick the simplest model that captures your seasonality: exponential smoothing as the baseline to beat, ARIMA/SARIMA for short well-behaved series, Prophet when you have multiple seasonalities, holidays and gaps and want sane defaults. Gotchas: evaluate on rolling-origin splits, never a random split; a forecast is a distribution, so report intervals; and check residuals per season, not just overall.

Reference: Prophet.

LINKS IN THIS NOTE

[[ GAMs & LASSO ]] A GAM fits one smooth curve per feature and adds them up - flexible shape, still readable effects; it was my thesis workhorse with leading indicators. LASSO pulls the other lever: keep the model linear and let the penalty push...