Vancouver — 6-month directional accuracy
75.0%
Always-DOWN baseline: 67.4%
| Edge: +7.6 pp
Honest-edge metrics
| Metric | Value | Context |
|---|---|---|
| Directional accuracy | 75.0% | What the model gets right |
| Always-DOWN baseline | 67.4% | Trivial in this regime |
| Edge over baseline | +7.6 pp | The honest number |
| Balanced accuracy | 68.4% | (UP recall + DOWN recall) / 2 |
| UP precision | 65.6% | When model says UP, how often it's right |
| UP recall | 49.3% | Of actual UPs, how many caught |
| RMSE | 0.0256 | From trend_metrics.json |
Rolling-12-month accuracy vs. always-DOWN baseline
The shaded region is the model's edge over the trivial classifier in each rolling window. If the blue line slides toward (or below) the dashed baseline in recent months, the regime is shifting and the model's edge is collapsing.
Per-month breakdown
One row per prediction month. Color reflects edge over the always-DOWN baseline within that month: red = negative edge, yellow = 0–5 pp, green = > 5 pp.
| Month | Cells | Actual UP | Actual DOWN | DOWN-baseline | Model | Edge | UP correct / pred | DOWN correct / pred |
|---|---|---|---|---|---|---|---|---|
| 2024-01 | 77 | 73 | 4 | 5.2% | 94.8% | +89.6 pp | 73 / 77 | 0 / 0 |
| 2024-02 | 77 | 68 | 9 | 11.7% | 88.3% | +76.6 pp | 68 / 77 | 0 / 0 |
| 2024-03 | 77 | 46 | 31 | 40.3% | 40.3% | +0.0 pp | 0 / 0 | 31 / 77 |
| 2024-04 | 77 | 23 | 54 | 70.1% | 70.1% | +0.0 pp | 0 / 0 | 54 / 77 |
| 2024-05 | 77 | 7 | 70 | 90.9% | 90.9% | +0.0 pp | 0 / 0 | 70 / 77 |
| 2024-06 | 77 | 9 | 68 | 88.3% | 88.3% | +0.0 pp | 0 / 0 | 68 / 77 |
| 2024-07 | 77 | 19 | 58 | 75.3% | 75.3% | +0.0 pp | 0 / 0 | 58 / 77 |
| 2024-08 | 77 | 17 | 60 | 77.9% | 77.9% | +0.0 pp | 0 / 0 | 60 / 77 |
| 2024-09 | 77 | 35 | 42 | 54.5% | 54.5% | +0.0 pp | 0 / 0 | 42 / 77 |
| 2024-10 | 77 | 32 | 45 | 58.4% | 58.4% | +0.0 pp | 0 / 0 | 45 / 77 |
| 2024-11 | 77 | 52 | 25 | 32.5% | 67.5% | +35.1 pp | 52 / 77 | 0 / 0 |
| 2024-12 | 77 | 45 | 32 | 41.6% | 58.4% | +16.9 pp | 45 / 77 | 0 / 0 |
| 2025-01 | 77 | 33 | 44 | 57.1% | 42.9% | -14.3 pp | 33 / 77 | 0 / 0 |
| 2025-02 | 77 | 23 | 54 | 70.1% | 46.8% | -23.4 pp | 19 / 56 | 17 / 21 |
| 2025-03 | 77 | 18 | 59 | 76.6% | 76.6% | +0.0 pp | 0 / 0 | 59 / 77 |
| 2025-04 | 77 | 11 | 66 | 85.7% | 85.7% | +0.0 pp | 0 / 0 | 66 / 77 |
| 2025-05 | 77 | 6 | 71 | 92.2% | 92.2% | +0.0 pp | 0 / 0 | 71 / 77 |
| 2025-06 | 77 | 5 | 72 | 93.5% | 93.5% | +0.0 pp | 0 / 0 | 72 / 77 |
| 2025-07 | 77 | 4 | 73 | 94.8% | 94.8% | +0.0 pp | 0 / 0 | 73 / 77 |
| 2025-08 | 77 | 9 | 68 | 88.3% | 88.3% | +0.0 pp | 0 / 0 | 68 / 77 |
| 2025-09 | 77 | 14 | 63 | 81.8% | 81.8% | +0.0 pp | 0 / 0 | 63 / 77 |
| 2025-10 | 77 | 15 | 62 | 80.5% | 80.5% | +0.0 pp | 0 / 0 | 62 / 77 |
| 2025-11 | 77 | 13 | 64 | 83.1% | 83.1% | +0.0 pp | 0 / 0 | 64 / 77 |
| 2025-12 | 77 | 26 | 51 | 66.2% | 68.8% | +2.6 pp | 7 / 12 | 46 / 65 |
Per property-type breakdown
| Property type | Cells | Within-PT majority baseline | Model accuracy | Edge | RMSE |
|---|---|---|---|---|---|
| Apartment | 456 | 71.3% | 77.2% | +5.9 pp | 0.0308 |
| Composite | 480 | 69.0% | 77.1% | +8.1 pp | 0.0195 |
| Detached | 480 | 65.2% | 71.7% | +6.5 pp | 0.0264 |
| Townhouse | 432 | 63.9% | 74.1% | +10.2 pp | 0.0248 |
"Within-PT majority baseline" = max(P(UP), P(DOWN)) of actuals restricted to that property type. This is the trivial classifier specific to the subgroup; it differs from the overall always-DOWN baseline because UP/DOWN balance varies by PT.
Confusion matrix
Per-area breakdown — top 10 best, bottom 10 worst
Aggregated across property types within each area. Areas with < 12 cells excluded as small-sample noise.
Top 10 — highest accuracy
| Area | Cells | Baseline | Model | Edge |
|---|---|---|---|---|
| North Vancouver | 96 | 63.5% | 87.5% | +24.0 pp |
| Port Coquitlam | 96 | 69.8% | 86.5% | +16.7 pp |
| Coquitlam | 96 | 75.0% | 86.5% | +11.5 pp |
| Vancouver East | 96 | 71.9% | 81.2% | +9.4 pp |
| Burnaby East | 96 | 74.0% | 81.2% | +7.3 pp |
| Maple Ridge | 96 | 74.0% | 80.2% | +6.2 pp |
| West Vancouver | 72 | 73.6% | 79.2% | +5.6 pp |
| Burnaby North | 96 | 64.6% | 78.1% | +13.5 pp |
| Richmond | 96 | 81.2% | 78.1% | -3.1 pp |
| Vancouver West | 96 | 69.8% | 76.0% | +6.2 pp |
Bottom 10 — lowest accuracy
| Area | Cells | Baseline | Model | Edge |
|---|---|---|---|---|
| Bowen Island | 48 | 58.3% | 33.3% | -25.0 pp |
| Whistler | 96 | 52.1% | 64.6% | +12.5 pp |
| Sunshine Coast | 96 | 63.5% | 65.6% | +2.1 pp |
| Squamish | 96 | 53.1% | 67.7% | +14.6 pp |
| Pitt Meadows | 96 | 67.7% | 68.8% | +1.0 pp |
| Tsawwassen | 96 | 63.5% | 69.8% | +6.2 pp |
| Ladner | 96 | 59.4% | 71.9% | +12.5 pp |
| Burnaby South | 96 | 69.8% | 72.9% | +3.1 pp |
| Port Moody | 96 | 70.8% | 75.0% | +4.2 pp |
| New Westminster | 96 | 77.1% | 76.0% | -1.0 pp |
Live forecasts (un-verified)
Forecast made 2026-04.
Verifiable starting 2026-11
(when 2026-10 HPI lands).
81 (area, property_type) cohort predictions in
forecast_202604_6m.csv.
Top 10 by predicted return
| Area | Property type | Benchmark price | Pred return | Direction | Future price | $ change |
|---|---|---|---|---|---|---|
| Whistler | Apartment | $515,100 | −1.49% | DOWN | $507,411 | −$7,689 |
| Squamish | Apartment | $491,900 | −1.56% | DOWN | $484,222 | −$7,678 |
| West Vancouver | Apartment | $1,017,000 | −1.59% | DOWN | $1,000,823 | −$16,177 |
| Sunshine Coast | Apartment | $389,600 | −1.60% | DOWN | $383,363 | −$6,237 |
| Sunshine Coast | Composite | $768,000 | −1.64% | DOWN | $755,408 | −$12,592 |
| Sunshine Coast | Detached | $820,700 | −1.64% | DOWN | $807,244 | −$13,456 |
| Whistler | Composite | $1,290,500 | −1.67% | DOWN | $1,268,921 | −$21,579 |
| Maple Ridge | Townhouse | $723,900 | −1.68% | DOWN | $711,744 | −$12,156 |
| Port Coquitlam | Apartment | $580,400 | −1.68% | DOWN | $570,647 | −$9,753 |
| Vancouver East | Detached | $1,681,000 | −1.68% | DOWN | $1,652,717 | −$28,283 |
Bottom 10 by predicted return
| Area | Property type | Benchmark price | Pred return | Direction | Future price | $ change |
|---|---|---|---|---|---|---|
| Port Moody | Apartment | $703,400 | −1.95% | DOWN | $689,667 | −$13,733 |
| Burnaby North | Townhouse | $931,900 | −1.95% | DOWN | $913,706 | −$18,194 |
| Tsawwassen | Detached | $1,530,000 | −1.95% | DOWN | $1,500,130 | −$29,870 |
| Vancouver East | Apartment | $664,800 | −1.95% | DOWN | $651,821 | −$12,979 |
| Ladner | Townhouse | $983,600 | −1.95% | DOWN | $964,397 | −$19,203 |
| Port Coquitlam | Detached | $1,326,700 | −1.95% | DOWN | $1,300,828 | −$25,872 |
| North Vancouver | Townhouse | $1,273,700 | −1.95% | DOWN | $1,248,862 | −$24,838 |
| North Vancouver | Detached | $2,129,900 | −1.95% | DOWN | $2,088,365 | −$41,535 |
| North Vancouver | Apartment | $784,500 | −1.95% | DOWN | $769,201 | −$15,299 |
| North Vancouver | Composite | $1,320,500 | −1.95% | DOWN | $1,294,749 | −$25,751 |
Track record (forward-prediction history)
Forecast archive started 2026-05. The first scoreable forecast
becomes verifiable ~6 months after it was made (once the target
month's HPI is published). Historical pre-archive months are not
recoverable; the per-month backtest table above is the available
proxy until archive matures.