Decision guide

Choosing a forecasting algorithm

Clarex has two forecasting algorithms — LightGBM forecasting and AutoARIMA. Both take a single time series and project it forward over a horizon, and both return a point forecast plus an uncertainty band. They come at the problem from opposite traditions: one is classical statistics, the other is modern gradient boosting. This guide helps you pick — and the honest answer is often run both.

At a glance

	AutoARIMA	LightGBM forecasting
Approach	Classical statistics (SARIMAX)	Gradient-boosted decision trees
Best fit	Regular, stable, seasonal series	Non-linear structure, many interacting drivers
Uncertainty band	Analytic — derived from the model	Quantile regression, conformally calibrated
Exogenous regressors	Yes	Yes (as engineered features)
Configuration	Automatic model search	Some tuning

Start with AutoARIMA as the yardstick

AutoARIMA is the classical baseline — the model every forecaster is expected to run as a reference point. It searches for the best ARIMA structure automatically, then produces a forecast with a statistically derived prediction interval for every period. It is at its best when the series is regular and reasonably stable — steady seasonality, no wild regime changes.

Run it first: it is fast, it needs almost no configuration, and its honest, model-derived intervals are a trustworthy sanity check on anything more complex.

When LightGBM forecasting wins

LightGBM forecasting treats forecasting as a supervised-learning problem: it engineers calendar and lag features and trains a gradient-boosted model on them. That lets it capture patterns a classical ARIMA model struggles with — non-linear effects, interactions between drivers, irregular structure — and it scales comfortably from tens of rows to millions. Its uncertainty band comes from quantile regression with a conformal calibration step, so the band stays honest even when the model is complex.

It can also project genuinely beyond the end of your data — the future-forecast analysis, currently a LightGBM-only capability in Clarex.

Reach for it when your series has structure ARIMA can't express, when you have several exogenous drivers that interact, or when AutoARIMA's accuracy simply isn't good enough.

The honest recommendation: run both

Forecasting is one of the few places where it costs almost nothing to try both and compare. They are cheap to run, they report comparable error metrics, and the comparison itself is informative: a classical model and a gradient-boosted model failing in the same way is strong evidence the signal is genuinely hard; disagreeing tells you something too. Use AutoARIMA as the always-on baseline, and keep the LightGBM forecast whenever it beats it on your hold-out metric.

Rules of thumb

Regular, stable, seasonal series? AutoARIMA is often all you need.
Non-linear structure, interacting drivers, or lots of history? LightGBM forecasting.
Want trustworthy intervals from first principles? AutoARIMA's are analytic.
Need to forecast past the end of the data? LightGBM forecasting.
Not sure? Run both — the comparison is the point.
Predicting from feature columns with no time order? This isn't a forecasting problem — see Choosing a regression algorithm (numeric target) or Choosing a classification algorithm (category target).