Examples and validation scope

Run all commands from the skill root in the isolated TimesFM environment. Choose output directories outside the shipped skill tree. Plots require Matplotlib; GIF generation also requires Pillow.

python examples/global-temperature/run_forecast.py --output-dir /tmp/timesfm-temperature
python examples/global-temperature/visualize_forecast.py --output-dir /tmp/timesfm-temperature
python examples/global-temperature/generate_animation_data.py --output-dir /tmp/timesfm-temperature
python examples/global-temperature/generate_gif.py --output-dir /tmp/timesfm-temperature
python examples/global-temperature/generate_html.py --output-dir /tmp/timesfm-temperature
python examples/anomaly-detection/detect_anomalies.py --output-dir /tmp/timesfm-anomalies
python examples/covariates-forecasting/demo_covariates.py --output-dir /tmp/timesfm-covariates

Global-temperature and anomaly scripts download/use the pinned 2.5 checkpoint and run preflight. The bundled temperature-shaped values have unverified provenance; they are illustrative input, not a current NASA/NOAA data release. Retired generated results were removed instead of relabeling incorrect bands. The corrected schema rejects legacy results in plot/animation tools.

The covariate demo generates 108 synthetic rows (3 stores x 36 weeks) and plots known generator components. It does not execute pretrained forecasting or establish that adding covariates improves a model. Its 24-point context supports the xreg + timesfm example; timesfm + xreg requires a longer history (>32).

Checks before reporting a result

• Check model/package/revision, cadence, target units and forecast origins.
• Retain positions for missing timestamps/values; fit transformations per origin.
• Check all forecast values are finite, output dimensions match, and quantiles are ordered. For 2.5 point=q[...,5], while 3.0 point=q[...,4].
• Report nominal interval level correctly: q10-q90 is 80%, q20-q80 is 60%.
• Evaluate against held-out actuals and naive baselines; report coverage and width by horizon, not only mean accuracy.
• Qualify illustrative retrospective screens and synthetic injected events; observed flags do not validate alarm severity or clinical/climate meaning.
• Plot every band's extent and align each successive forecast with its own historical cutoff. A final forecast is only a retrospective reference layer.

Executed checks for this refresh

The repository suite covers strict CSV/date handling, missingness, quantile indices/shapes, system checks, native 2.5 preprocessing/compile/XReg with a controlled lightweight decoder, tiny random 2.5 native transformer/decode, and tiny random 3.0 native save/load/inference. Example plots and animations are tested with clearly synthetic model outputs. These verify API mechanics and file/visual alignment, not pretrained accuracy, calibration, cloud availability or GPU/MLX performance. Checkpoint-load snippets remain illustrative until run against the real weights for the user's data.

From the repository root:

uv run skills-ref validate skills/timesfm-forecasting
python tests/run_all.py --isolated timesfm-forecasting