Forecast intervals
Choose Observed bars in Forecast. The interval controls the historical series and each future forecast step; it does not merely change the chart display. The same selection is retained when reopening, refreshing or downloading a forecast.
Use
1Month for months and lowercase 1m for minutes. Calendar weeks begin Monday at 00:00 UTC; months begin on the first at 00:00 UTC. A monthly forecast step is a real calendar month, including February and leap years, rather than 30 elapsed days.
Completed observations
Stocks and Dukascopy weekly/monthly bars aggregate genuine daily observations. Stock daily bars are grouped by the provider’s exchange session date; Dukascopy daily bars use UTC dates. Four-hour stock bars aggregate genuine hourly observations. Prediction markets and perpetuals aggregate their available native observations. Empty buckets remain absent; open buckets are excluded. No observation is invented to meet a model’s minimum history requirement. Download rows retain bucket-start timestamps where the provider uses them. The forecasting service converts those closes to their availability at the bucket end. Kalshi end-stamped candles remain end-stamped. Results identify the effective provider and input cutoff. Daily US-stocktrading_sessions horizons use NYSE sessions and exclude exchange holidays. Other market intervals use frequency_periods and a UTC calendar grid: a predicted timestamp can fall during a market closure and is not evidence of an executable quote. Custom CSV offsets retain the existing dataset workflow.
Submit a forecast
prediction_length counts bars. For example, 12 steps at 5 minutes forecast one hour; 3 steps at 4 hours forecast 12 hours; 3 monthly steps forecast three calendar boundaries. The website offers duration units and an exact calendar end time.
GET /api/v1/forecast/models before choosing models. It returns a frequencies array alongside model capabilities. Model access, minimum context, requested quantiles and horizon limits still apply. An hourly sports market may have enough history for an ensemble while having too few weekly/monthly observations; the service reports that failure instead of fabricating a history.
For an exact end instant use prediction_end_at, with a timezone, and include source.frequency. The server counts only future period boundaries through that instant. An end before the next complete period cannot create a forecast step.
Download the same interval
POST /api/market-data/stocks/history accepts the market aliases above in timeframe. Dukascopy additionally supports paged downloads and separate bid/ask closes. Follow every next_cursor with unchanged settings before exporting the combined snapshot. Weekly pages retain weeks crossing annual archive boundaries.
Event exports use 1m, 5m, 15m, 30m, 1h, 4h, 1d, 1w or 1month. Perpetual history accepts the canonical forecast frequencies. Provider retention, request bounds, entitlements and rate limits remain in effect; selecting an interval does not guarantee 500 observations.
Forecast API · Market downloads · MCP access