Time series and forecasting · insight-1252
Match a foundation model's context window to your longest seasonal cycle
- Evidence type
- empirical finding
- Difficulty
- medium
- Roles
- Data scientist, ML engineer, Applied AI
- Industries
- energy, telecom, logistics
Summary: situation, action, reason, boundary
Situation
Hourly electricity demand has weekly and annual cycles, and a pretrained forecaster gets the last 512 hours as context.
Action
Compute the context needed (longest cycle times at least two repeats) and compare it with the model's supported length. If short, forecast a seasonally adjusted series, add calendar covariates, or use a longer-context model, and backtest context length.
Reason
Chronos was trained with a context of 512, and its ablation found gains up to 1024 before saturation. TimesFM notes a rise from 2048 to 16k, so supported lengths differ by model and version.
Boundary
High-frequency data with long cycles such as hourly load, minute-level traffic or daily series with annual seasonality.
Symptom or risk
The model sees about three weeks and not last year's level at the same date, so annual seasonality is silently absent.
Exceptions
If a seasonally adjusted or calendar-augmented series is supplied, a short window may be enough.
Trade-offs
Longer contexts raise inference cost and did not always improve accuracy in the Chronos ablation.
Pitfalls
Assuming that a longer context window in the API means the model uses it well.
Example
Hypothetical: two annual cycles of hourly data are 2 x 8,760 = 17,520 observations, about 34 times a 512-step context.
Lessons and concepts
Claims and their sources
Chronos was trained with a context length of 512, and its ablation reports improvement with longer training context up to 1024 and saturation or decline beyond.
As read on 2026-10-12, the TimesFM README says TimesFM 2.5 (15 Sept 2025) supports up to 16k context length, up from 2048; the 3.0 section gives no maximum.