Input Features (per unit)

• Complaints (Stage 1) per 1,000 units — CH01_1 / units × 1000
• ASB complaints per 1,000 units — CH02_1 / units × 1000
• Arrears rate — current_arrears / rents × 100
• Routine maintenance per unit — routine_maintenance / units
• Planned maintenance per unit — planned_maintenance / units
• Void rate — void_losses / rents × 100
• Operating margin — operating_surplus_deficit / turnover × 100
• Turnover per unit — turnover / units

Step 1 — Correlation

For each feature, we calculate the Pearson correlation between that feature and the actual year-over-year TP01 change (current year minus previous year) across all organisations with two years of data.

Step 2 — Per-Feature Regression

Each feature is normalised to 0–1 across all orgs. A linear regression is fitted per feature: predicted_change = slope × normalised_feature + intercept.

Step 3 — Weighted Ensemble

The top 3 most correlated features are combined. Each feature's prediction is weighted by its correlation strength squared: weight = |correlation|². The final delta is the weighted average.

Step 4 — Calibration

Raw deltas are z-score normalised, then rescaled to match the observed mean and standard deviation of actual TP01 changes. This ensures predictions reflect real-world volatility.

Band Assignment

• Improving — predicted change > +1.0 percentage point
• Stable — predicted change between −1.0 and +1.0
• Declining — predicted change < −1.0 percentage point