Forecast

Upload a CSV with a header row, a date/time column and one numeric target. Select both columns after upload. Use 40–10,000 rows, at most 2 MB. Values must be finite; missing values are rejected, never invented.

timestamp,value
2026-09-01T09:30:00-04:00,100.5
2026-09-01T09:31:00-04:00,101.2

ISO dates represent daily observations. Date-times without an offset use your device timezone; include an explicit offset around daylight-saving transitions. Rows are sorted and duplicate timestamps use the last submitted value.

The selected columns are sent securely for the forecast job. Guest results belong to your guest session; an account is needed to keep requests in your profile. Uploaded data is not automatically public.

Last N observations

Maximum actual input bars for ticker and event forecasts. The cutoff and game-history filters apply first, then the latest N valid observations are selected. N bars can span more than N minutes if history has gaps. Each enabled model still enforces its own minimum history.

Context length

Maximum recent observations passed to foundation models within that selected input. The worker uses the available history and each model's limits; it never invents missing bars. Prophet can use the full permitted history.

Target transform

Auto uses logarithms for strictly positive values; None uses original values. Log requires positive inputs. Prediction-market probabilities use the service's bounded probability method automatically.

Model failure policy

Fail reproducibly stops if a selected model fails. Renormalize continues only if a surviving model supports every requested quantile, and reports the actual participating models and weights.

Dataset interval

Only needed for uploaded datasets. Detect reads timestamp spacing; choose an explicit interval for irregular data. Market data already has an interval selector, so no second setting is needed. Custom intervals use supported pandas frequency notation, such as 2h.

Data cutoff and forecast end

Choose an exact calendar date/time in your device timezone, or a duration in days, hours or minutes. Historical cutoffs can be 120 or 180 days ago when the provider retains the requested bars. Latest uses the newest available observation. A 30-minute cutoff plus a 60-minute forecast overlays already-observed outcomes separately. This is a replay generated now, not a forecast published earlier.

Game history

Auto uses pregame plus in-game history until 32 elapsed game minutes, then only in-game observations. Missing game start keeps both phases. Flat genuine observations are allowed; gaps are never filled.

Toto size

Choose an approved checkpoint from 4M through 2.5B. Larger checkpoints need more memory and startup time. The chosen revision is stored with your request.

Game history and lookback

Choose pregame, in-game, or both using the provider's recorded start. Set the history window to 60 minutes for only the last hour; zero uses up to 500 available bars. Genuine unchanged quotes can be forecast with a low-information warning. Missing prices are never filled and artificial variation is never added.

Quantiles and weights

P10 and P90 mark lower and upper positions in the model's forecast distribution—not guaranteed price boundaries. Equal entered weights are normalized per quantile; Toto and TimesFM do not supply P1 or P99.