alterlab-timesfm

TimesFM Forecasting

Overview

TimesFM (Time Series Foundation Model) is a pretrained decoder-only foundation model developed by Google Research for time-series forecasting. It works zero-shot — feed it a time series and it returns point forecasts with quantile prediction intervals, no training required. The timesfm package (current 3.0.2) ships two model APIs:

• TimesFM 2.5 (200M, Apache-2.0 weights) — timesfm.TimesFM_2p5_200M_torch with ForecastConfig; univariate, optional covariates via XReg. The default in this skill and the only choice for commercial or production use.
• TimesFM 3.0 (~330M, released Aug 2026) — timesfm3.TimesFM3Forecaster; univariate and multivariate forecasting with native past-only and past-and-future covariates. Its weights are under timesfm-non-commercial-license-v1.0, so use it only for non-commercial research and tell the user about the restriction.

This skill wraps TimesFM for safe, agent-friendly local inference. It includes a mandatory preflight system checker that verifies RAM, GPU memory, and disk space before the model is ever loaded so the agent never crashes a user's machine.

Key numbers: TimesFM 2.5 uses 200M parameters (0.93 GB safetensors); TimesFM 3.0 uses ~330M (1.3 GB). Run the system checker before the first load so an under-resourced machine fails fast instead of swapping or crashing.

When to Use This Skill

Use this skill when:

• Forecasting any univariate time series (sales, demand, sensor, vitals, price, weather)
• You need zero-shot forecasting without training a custom model
• You want probabilistic forecasts with calibrated prediction intervals (quantiles)
• You have time series of any length (the model handles 1–16,384 context points)
• You need to batch-forecast hundreds or thousands of series efficiently
• You want a foundation model approach instead of hand-tuning ARIMA/ETS parameters
• You have related channels or known future drivers (TimesFM 3.0 multivariate + covariates, or TimesFM 2.5 XReg)