Problem
Teams that forecast business metrics usually train one model per series or group, which is slow to run and monitor. Metrics with weekly, monthly or quarterly frequency often have about 12 points or fewer. Large companies have the opposite problem: many interconnected series (promotions, ad campaigns, service metrics). Papers on pretrained models are hard to compare, so it is unclear whether one can replace the zoo.
Example from the article: on the speaker's team's internal datasets, zero-shot Chronos was the best foundation model. On sMAPE and the 75th percentile of errors, seasonal naive was still best, because pretrained models leave heavy error tails.
Solution in brief
Use a pretrained time-series model (TimesFM, Chronos, Moirai, Lag-Llama) as a zero-shot baseline: one API call returns a forecast with no per-series training. Compare it on your own data against seasonal naive, a classical ensemble and boosting, and fine-tune if the gap matters. Zero-shot models are not yet state of the art but are worth trying.
Questions this case answers
- How to forecast many short business metric series with a pretrained sequence model in zero-shot mode and check it against a seasonal naive baseline on one's own data
- How to forecast many interrelated series that are driven by campaign or promotion regressors with a pretrained model
- How to forecast business metrics with a pretrained time-series foundation model, without training one per series
- How to choose between TimesFM, Chronos, Moirai, Lag-Llama and a classical baseline
Tools and methods
- TimesFM (Google) — decoder-only, patches, about 200M parameters
- Chronos (Amazon) — T5 on quantized values, 700M parameters
- Moirai / Uni2TS (Salesforce) — patches, Any-Variate Attention
- Lag-Llama — Llama with lags as tokens, Student's t output, open weights
- TimeGPT-1 (Nixtla) — closed model, paid API only
- Baselines — seasonal naive, AutoARIMA/AutoETS/AutoCES/Theta ensemble, CatBoost
Google or a one-time email code. No password.
