Roadmap: predictive occupancy — P(occupied within N minutes) and likely-next-area

[Hankanman]

The killer version

Move from detecting occupancy to predicting it: a P(occupied within 10 min) sensor per area and a likely_next_area attribute, driven by the 168-bucket time priors and the learned 2-hop transition chains that adjacent-areas (#454) already records. Pre-heat the bathroom because the household always showers at 7am after leaving the bedroom; light the hallway before the door opens.

This project's asset

The hard data already exists on every install running adjacency: AreaTransitions chains with hour-of-week bucketing and six-level smoothing fallback. No new learning infrastructure — this is a read-side feature over Phase-3 data.

First three steps (the first one costs nothing)

1. Offline evaluation: from an existing database, replay history — for each area-entry event, would the preceding chain + hour bucket have predicted it? Measure precision/recall at 5/10/15-minute horizons
2. Pick a publishable operating point (e.g. precision ≥ 0.7 at whatever recall the data supports) — if the data can't support one, the epic stops here and we say so
3. Ship the prediction sensor (disabled by default) with the trust score (Roadmap: per-area trust score — measured calibration, false-transition rate, and earned auto-threshold #499) extended to score predictions the same way it scores detections

You have a result when…

The offline evaluation clears the chosen bar on two real homes, and a shipped prediction demonstrably fires before physical entry in live use. This is also the flagship write-up candidate — a measured next-room-prediction result from purely local learning is genuinely novel in the HA ecosystem.

Depends on

Adjacency data maturity (feature shipped 2026-07-06 in 2026.7.1-pre1) and #499 for scoring.

[coderabbitai]

🔗 Related PRs

#357 - Add purpose-aware activity detection sensors [merged]
#454 - feat: adjacent-areas (next-door room influence) [merged]
#456 - feat: adjacent-areas Phase 4 — Bayesian boost + decay modifier (rolls into #454) [merged]

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!