Predicting Neighborhood Affordability Tiers in Toronto Using Static Urban Features

Abstract

We present a machine learning system for predict- ing which affordability tier (Budget, Moderate, Expensive, or Premium) a Toronto neighborhood will occupy two years ahead, using only non-rent urban characteristics. Our dataset spans 15 years (2010–2024) across 158 neighborhoods, integrating rental, crime, transit, and green space data. We conducted a randomized hyperparameter search across 60 configurations of three model families evaluated under both stratified and temporal cross- validation. The best model (Gradient Boosting) achieves a cross- validated macro F1 of 0.740 and a held-out test F1 of 0.604 (95% CI: 0.560–0.644), a 2.4× improvement over random baseline. The convergence of all three model families to similar performance provides evidence that the ceiling is data-driven. We complement the classifier with K-Means clustering and deploy the system as an interactive web application. Code: github.com/Western-Artificial- Intelligence/condo-cost-predictor.