Zero-shot time-series forecasting with Google's TimesFM 3.0 foundation model. Give it a history, get a quantile forecast — no training, optional covariates. Weights are under the TimesFM Non-Commercial License v1.0.
Signature | ||||
|---|---|---|---|---|
JSON→JSON | Public | 12 |
Usage
12 callsUpgraded you to Gold!
One blocker needs your call: the forecast-ledger dataset can't be created — create_dataset fails with "Dataset limit reached (82/20)". The account holds 69 datasets, almost all stale one-off artifacts from the old weekly workflow (per-week copper/gold/silver/oil forecast snapshots from 2024–mid-2026, plus ~15 duplicate yfinance-btc-usd-* and fred-cbbtcusd-* downloads). The new program keeps everything in a single ledger, so those are dead weight.
Two ways to unblock:
I delete ~60 of the stale datasets (I'll list exact IDs before deleting; nothing under the new program is touched). No cost, immediate.
Upgrade the plan to Gold for a higher limit.
The 24 forecast rows already issued (UNRATE + PCOPPUSDM, 12 steps each, with route action receipts 01a09094-386a… and 01a09094-47a1…) are safe in scratch and will be written the moment a slot frees. Which way do you want to go?
Confirmed — the Gold upgrade lifted the dataset cap, and the ledger now exists.
forecast-ledger is live in the forecasting team, public: 24 forecast rows ingested (12 monthly steps each for UNRATE and PCOPPUSDM, issued 2026-09-11 before any target was knowable), every row carrying its Forecast route action receipt as a native reference. Schema has the quantiles, the seasonal-naive baseline, and the scoring columns (realized_value, abs_error, inside_80, status enum).
So the full spine is unblocked: TimesFM and Time Series Data routes work end-to-end, and forecasts are now being written to a queryable ledger before outcomes are known. First scoring opportunities: PCOPPUSDM step 1 (August copper, publishes ~mid-Sept) and UNRATE step 1 (September, after the Oct 2 release).