Empirical Valuation and Opportunity Detection in Emerging South Asian Real Estate: A Case Study of DHA Lahore Using Explainable AI (TreeSHAP)

Abstract

Automated Valuation Models (AVMs) are transforming property underwriting and investment benchmarking, yet their application to non-disclosure emerging real estate markets in South Asia remains critically under-researched. In master-planned urban developments such as the Defence Housing Authority (DHA) Lahore, plot valuations exhibit steep non-linear volatility across phases, blocks, commercial versus residential zoning, and micro-locational amenities that parametric linear appraisals fail to capture. This paper presents a leak-free empirical machine learning benchmark evaluating Ordinary Least Squares (OLS) Multiple Linear Regression against tree-based ensembles (Random Forest, XGBoost, and LightGBM) across N = 244 curated plot listings (200 residential and 44 commercial CCA plots) spanning DHA Lahore Phases 2 through 9 Prism. Incorporating out-of-fold hierarchical Bayesian target encoding of per-Marla block rates alongside zoning, road frontage, and text-mined possession features, Random Forest achieved the highest 5-fold stratified cross-validation stability (R² = 0.5806), while XGBoost achieved top hold-out accuracy with a Test R² of 0.8462, RMSE of 1.6566 Crore PKR, MAE of 0.9887 Crore PKR, and MAPE of 19.56% (reducing percentage valuation error by over half relative to OLS at 41.74%). Furthermore, we deploy exact polynomial-time TreeSHAP to decompose rupee-level feature contributions and construct a residual opportunity-detection pipeline (Δᵢ ≤ −15%) that isolates undervalued distressed listings (23.4%) with transparent, auditable justifications.