Auto-selection & conformal intervals

Auto-selection

You don’t have to pick a model at all: auto is the default. A call with no model option runs a rolling-origin backtest of each candidate on the series and forecasts with the lowest-error one. The choice is deterministic: the same rows pick the same winner.

SELECT forecast(ts, value, 12) FROM readings;   -- auto, implicitly

Selection backtests every candidate per call, which costs single-digit milliseconds on typical series and tens of milliseconds at the 4096-point context cap. On a latency-critical path over long series, pin a model or lower context_limit.

The default pool is the bundled statistical models (theta-classic, stub-seasonal-naive, tsb for intermittent demand) plus every eligible registered forecast student: distill a model once and auto competes it against the baselines automatically, no call-site changes. Eligibility is per call: a student trained for a shorter horizon than requested, or any student when interval_method is conformal, sits that call out quietly.

Pass candidates to narrow the pool explicitly, for example when many students are registered and you only want one backtested per call:

SELECT forecast(ts, value, 12,
  '{"model":"auto","candidates":["theta-classic","tsb","my-student"]}')
FROM readings;

The winning model’s id comes back in the result document’s model field, so you can see when your student beat the baselines.

Conformal intervals

The default prediction band is Gaussian, sized from the model’s in-sample error. On smooth data that band is overconfident. '{"interval_method":"conformal"}' replaces it with a distribution-free band calibrated on the model’s out-of-sample rolling residuals:

SELECT forecast(ts, value, 6, '{"interval_method":"conformal"}')
FROM readings;

On a smooth synthetic series the default band covered only 57% of points at a nominal 90% level; the conformal band landed at 90%. You can check coverage on your own data with backtest().

Conformal applies to the statistical models. A foundation-model student already emits its own quantile band, so asking for conformal on one is rejected rather than silently ignored. A series too short to calibrate returns insufficient_history instead of a bogus interval.