forecasting

name	forecasting
description	The output contract for producing a structured probabilistic forecast — the JSON shape, the calibration and quantile rules, and how to submit it. Load this ONLY when your task payload asks for a forecast; ignore it for open-ended questions. No scripts.

Forecasting skill

Load this when your task payload asks for a structured forecast. For open-ended questions, ignore it and just answer.

What you'll receive

A JSON payload describing the task: a task id, the as_of cutoff date, horizons (steps ahead), the standard_quantiles grid, a target_summary, the recent target_history_csv, and an output_schema showing the exact JSON to return.

The output contract

1. Produce one forecast per horizon in horizons.
2. Use exactly the levels in standard_quantiles — no additions or omissions.
3. point_forecast must equal the 0.50 quantile value.
4. Quantile values must be non-decreasing as the quantile level rises.
5. Use ONLY information available on or before as_of.
6. Put your reasoning in the rationale fields.

Submit by calling set_model_response with a json_response string that matches the payload's output_schema exactly — use "horizon" (an integer), and make "quantiles" a list of {"quantile": <level>, "value": <number>} objects. Omit any field not shown in the schema.

Calibration

Report calibrated intervals, not false precision: across many forecasts where your 80% band is stated, the truth should land inside it about 80% of the time. Anchor the point on the recent level and trend; let recent volatility set how wide the bands are, and widen them as the horizon grows.

Domain focus (edit this for your use case)

For monthly food CPI, anchor on the last index value and its recent year-over-year change; the index moves slowly and seasonally, so month-to-month jumps are small. Widen the bands with the horizon, and adjust for food-inflation or input-cost signals you have real evidence for — and say so in the rationale.

Room to grow

• Tighten the calibration guidance with your own backtest findings.
• Add worked examples of good vs. over-confident forecasts.