Compare forecasts with results
Forecast history lets you compare expectations with a game's later growth.
How to use it
- Open Forecasts in Analysis.
- Choose a game and forecast horizon.
- Compare the forecast and outcome for that horizon. The calculation method is available below.
Advice and practical context
Forecast as testable record
A forecast stores an expected range and horizon before the outcome arrives. Its value is that the original expectation cannot be silently rewritten after the fact. Compare it with the same metric and window used by the calculation.
Evaluate a series, not luck
One accurate forecast can be chance. Review error across several completed horizons, directional bias, and coverage. Do not include immature outcomes, cherry-pick successes, or turn past accuracy into a guarantee of the next result.
Use one error definition per horizon and keep games that declined; excluding failures inflates apparent quality. Split results when the method changes. Outcomes from different formula versions should not be presented as one continuous track record.
Technical detailsCalculations, permissions and behavior
Immutable forecasts and calibration
The first available daily forecast preserves raw features, percentiles, reasons, availability time and model version. Outcomes at 7/14/30/60/90 days compare the final week's mean with the initial daily mean. Breakout requires ≥3× and at least three days ≥2×; moderate growth ≥1.5×; stable 0.75–1.5×; decline <0.75×.
Incomplete outcomes remain unavailable. Calibration requires 200 unique games with 30-day outcomes, 30 breakouts, three genres and two age segments. Train/validation/test are chronological with overlapping horizons purged.
L2 logistic and Huber log-growth produce an inspectable candidate, never an automatically activated model. Only an administrator can explicitly apply, reject or roll back a version. CSV/JSON contains features, missingness, provenance and outcomes; export is limited to 10000 first game forecasts.