Forecasting model
Each region gets a next-24-hour, hourly forecast of its survivor and killer bonuses. This page explains what the forecast is and how to read it, at a conceptual level, not in code.
The forecast is available per region via the API (GET /api/v1/platforms/{platform}/regions/{region}/forecast) and is what the per-region page's forecast chart draws.
What it predicts
Bloodpoint incentives move with player population, which is strongly seasonal: a region tends to look similar at the same hour on the same kind of day from one week to the next (a Saturday evening resembles other Saturday evenings, a weekday morning resembles other weekday mornings).
The model exploits this. For each future hour it produces:
- a most-likely level (the median), and
- a p25–p75 band around it (the interquartile range), for each role.
Levels are quantized (bonuses come in discrete steps), and the model never averages those discrete levels into something that can't actually occur. It reports real levels and the spread between them.
How it's built
The forecast blends two ideas:
- Seasonal climatology: a day-of-week × hour-of-day picture of how the region usually behaves, computed in the region's local timezone (distinguishing weekday from weekend patterns). This captures the recurring weekly rhythm.
- Persistence: the recent, current value carries information about the very near future, so the blend leans on where the region is right now as well as where it usually is.
These are combined with recency weighting: more recent history counts for more than older history, so the forecast tracks gradual shifts in a region's behaviour rather than treating months-old data as equally relevant.
The history the model trains on is governed by FORECAST_WINDOW_DAYS (84 days by default), which is configured independently of the data retention window. See Configuration.
Confidence
Each forecast carries a confidence level (high, medium, or low) that reflects how much history backs it. A region with months of consistent readings yields a confident forecast. A region that's only recently gained coverage, or has sparse history, yields a low-confidence one. Use it to judge how much to trust the band.
Hindcast: the model graded against reality
The forecast chart also shows a hindcast (what the model would have predicted for the past day) drawn against the actual values that occurred.
The hindcast is leak-free: when computing the prediction for a past hour, the model only uses data it would have had before that hour, never the actual value it's being compared to. That makes the hindcast an honest, visual check of how well the model has been tracking reality recently. If the hindcast line hugs the actuals, the forward forecast is more trustworthy. If it diverges, treat the forecast with more caution.
The forecast response
The endpoint returns the region, platform, when it was generated, the horizon, the confidence level, and a list of hourly points, each with the median plus the low/high band for both roles. See the API Reference for the exact schema, and Using the API for where it sits among the other endpoints.