forecast-web
    Preparing search index...

    Forecasting

    Option Default Meaning
    model "timesfm-2.5" Which model to run. See MODELS.
    weights the model's published repo Hugging Face repo id, or a URL to a folder with the same layout for self-hosting.
    revision the commit this release was verified against Hugging Face revision. A custom weights repo defaults to "main".
    dtype "auto" Weight format: "auto", "q8", "fp16" or "fp32".
    device "auto" "auto", "webgpu" or "wasm".
    worker true Run inference in a Web Worker so the page stays responsive.
    cache true Keep downloads in the Cache API.
    wasmPaths jsDelivr Where onnxruntime-web fetches its .wasm files.
    onProgress none Called with { loaded, total, cached } during the download.
    signal none An AbortSignal that cancels the load.

    forecast(series, { horizon, calibrate }) accepts a horizon from 1 to the model's maxHorizon, defaults to the model's horizon (128 for TimesFM, 64 for Chronos-Bolt), and returns, per series, median and q10 through q90. The median is the point forecast. TimesFM also returns mean, its separate mean head. calibrate: true rescales the bands from a backtest; see Anomalies and calibrated bands.

    Calls on one model may overlap: they queue and each resolves with its own result. dispose() frees the model; every later call rejects with DISPOSED. clearCache() deletes downloaded model files, for one model or all of them:

    import { clearCache } from "forecast-web";

    await clearCache("timesfm-2.5"); // that model's published weights, every format
    await clearCache(); // everything, including self-hosted weights

    Both resolve true if anything was deleted.

    Every error is a ForecastError with a code. New codes may be added in minor releases, so handle unknown codes with a fallback.

    Code When
    UNKNOWN_MODEL model is not in MODELS.
    LOAD_FAILED Weights could not be downloaded, failed their SHA-256 check or do not match the model, or the backend could not start.
    ABORTED The load's signal fired.
    DISPOSED The model was disposed. Checked before anything else.
    EMPTY_INPUT No series were passed to forecast.
    INVALID_HORIZON The horizon is not a positive integer.
    HORIZON_TOO_LONG The horizon is above the model's maxHorizon.
    INVALID_INPUT Malformed input (including an empty table), an unknown sensitivity, or a series too short to check for anomalies.
    TOO_MANY_CLASSES A classification table has more than 10 classes.
    TOO_LARGE A table is over the size limits in LIMITS.