Open Doorway · Institutional Real Estate Intelligence
Model Accuracy Report, 2026-Q3
Headline results
Directional accuracy
99.8%
out-of-sample sign match
MAPE (12-month)
33.58%
549 of 549 eligible
R² (pooled)
0.484
across all folds
Predictions
549
11 walk-forward folds
Performance by market
| Market | Predictions | Directional acc. | MAPE |
|---|---|---|---|
| market_2 | 110 | 100.0% | 28.59% |
| market_1 | 110 | 99.1% | 31.73% |
| market_4 | 110 | 100.0% | 25.30% |
| market_0 | 110 | 100.0% | 27.48% |
| market_3 | 109 | 100.0% | 54.97% |
Fold detail
| Fold | Train end | Test window | Rows (train/test) | Directional acc. | MAPE |
|---|---|---|---|---|---|
| 0 | 2039-01-01T00:00:00 | 2040-01-01T00:00:00 → 2050-01-01T00:00:00 | 200 / 50 | 100.0% | 25.99% |
| 1 | 2049-01-01T00:00:00 | 2050-01-01T00:00:00 → 2060-01-01T00:00:00 | 250 / 50 | 100.0% | 26.66% |
| 2 | 2059-01-01T00:00:00 | 2060-01-01T00:00:00 → 2070-01-01T00:00:00 | 300 / 50 | 100.0% | 88.67% |
| 3 | 2069-01-01T00:00:00 | 2070-01-01T00:00:00 → 2080-01-01T00:00:00 | 350 / 50 | 98.0% | 35.40% |
| 4 | 2079-01-01T00:00:00 | 2080-01-01T00:00:00 → 2090-01-01T00:00:00 | 400 / 50 | 100.0% | 36.90% |
| 5 | 2089-01-01T00:00:00 | 2090-01-01T00:00:00 → 2100-01-01T00:00:00 | 450 / 50 | 100.0% | 35.88% |
| 6 | 2099-01-01T00:00:00 | 2100-01-01T00:00:00 → 2110-01-01T00:00:00 | 500 / 50 | 100.0% | 19.73% |
| 7 | 2109-01-01T00:00:00 | 2110-01-01T00:00:00 → 2120-01-01T00:00:00 | 550 / 50 | 100.0% | 26.32% |
| 8 | 2119-01-01T00:00:00 | 2120-01-01T00:00:00 → 2130-01-01T00:00:00 | 600 / 50 | 100.0% | 21.15% |
| 9 | 2129-01-01T00:00:00 | 2130-01-01T00:00:00 → 2140-01-01T00:00:00 | 650 / 50 | 100.0% | 26.18% |
| 10 | 2139-01-01T00:00:00 | 2140-01-01T00:00:00 → 2149-01-01T00:00:00 | 700 / 49 | 100.0% | 26.33% |
Methodology
Expanding-window walk-forward backtest with an embargo gap between the training window and each test fold (the embargo equals the 12-month target horizon at the dataset's sampling frequency), preventing information leakage from overlapping forecast horizons. A fresh model is fitted per fold using the production feature protocol. Reported metrics are computed only on out-of-sample predictions.
Minimum training rows per fold: 200 · Evaluation window and fold boundaries are listed in the fold table above.
Limitations & disclosures
- Directional accuracy is the share of out-of-sample predictions whose sign matched the realized 12-month price change.
- MAPE excludes observations with |actual| below 0.1 percentage points, where percentage error is unstable; the excluded count is disclosed in the headline block.
- beats_naive compares model MAPE against a persistence baseline (forecast = trailing annual appreciation). The model earns deployment confidence only while this flag stays True.
- Prediction-interval coverage is validated during model training but is not yet re-validated inside the walk-forward harness; treat P10-P90 band widths as indicative until then.