Time-series forecasting

Predicting the future values of a sequence from its own past, optionally with covariates. It is one of the three things ml-system-design observes actually ship in industry — alongside ranking and matching — and until timesfm-2-5-forecasting-tutorial the wiki had no page on it.

The task has a property the rest of this corpus’s subjects don’t: the answer arrives on its own. A classifier needs labels; a forecast is scored by waiting. That makes evaluation unusually tractable and unusually easy to fake, because the modeller controls where the “future” starts.

The evaluation apparatus, which is the transferable part

timesfm-2-5-forecasting-tutorial supplies the first full evaluation design in this wiki. Four pieces, none of them specific to the model being tested:

A baseline that isn’t zero. Seasonal-naive — repeat the last weekly cycle — is the standard reference, because a series with any periodicity makes an untrained repeat look good. Reporting an error without it says nothing. MASE (mean absolute scaled error) bakes the comparison into the metric: below 1 beats the naive forecast, above 1 loses to it, and the scale of the series cancels out. That last property is why MASE travels across series where MAPE breaks down (it explodes near zero and punishes over- and under-prediction asymmetrically).

Rolling-origin backtesting. One holdout window is one sample. Stepping the cut-off backwards through history — six folds of 56 days in the tutorial, aggregated over 36 store-fold combinations — turns a single number into a distribution and exposes whether the model only works on the last quarter. This is the forecasting name for what ../quant-trading-wiki’s backtesting does for trading strategies, and the failure mode both share is lookahead: any information from after the cut-off leaking into the model invalidates the result silently.

Scoring the interval, not just the line. A probabilistic forecast needs probabilistic scoring: pinball loss (quantile-weighted error, so being outside an interval costs more in the direction the quantile promised) and coverage (do 80% of actuals land inside the 80% band?). A model can have excellent point accuracy and dishonest intervals, and only these catch it.

Ablating the knobs. Sweeping context length across {64 … 1024} and reporting accuracy and inference cost turns a configuration flag into an informed decision — accuracy plateaus, cost doesn’t.

Two things a forecast is for

Prediction, and detecting that the world stopped matching the prediction. The second falls out of the first for free once the model emits intervals: compare the observation to the 80% band, grade it OK / WARNING / CRITICAL by how far outside it sits, and forecasting has become monitoring (timesfm-2-5-forecasting-tutorial). Anomaly detection here is not a separate model but a residual with a threshold.

The field around this page (2026-08-10)

awesome-agentic-time-series indexes 243 papers on LLMs and agents applied to time series, and its shape says something the individual entries don’t. Thirty-one foundation models are catalogued, so timesfm is one of a crowd rather than a novelty — Toto 2.0 announces “forecasting enters the scaling era,” Timer-S1 is billion-scale. But 16 papers cover forecast and decision evaluation against 133 on building systems. The section below asked for a comparison and the literature index confirms that the comparison is the field’s thinnest corner, not a gap in this wiki’s reading.

What this wiki cannot say yet

Whether a pretrained forecaster like timesfm beats the classical alternatives. The only source here compares against seasonal-naive on synthetic data it generated itself, with no ARIMA, no exponential smoothing, no gradient-boosted tabular baseline, and no real series anywhere. Those comparisons exist in the forecasting literature (the M-competitions are the standard venue) and none of it is in this corpus. Until a source with real data and real baselines lands, the wiki holds a method for evaluating forecasters and no result about any of them.

Related

timesfm · timesfm-2-5-forecasting-tutorial · ml-system-design · demo-to-production-gap · machine-learning · synthesis