A Stable and Interpretable NNLS Ensemble for House Price Prediction on the Ames Dataset

Abstract

Predicting house prices has broad real-world impact across valuation, lending, and taxation. This paper proposes a principled stacking method for the Ames Housing dataset that blends eight strong regressors—linear, bagging, and boosting—via a simplex-constrained Non-Negative Least Squares (NNLS) optimizer. By enforcing non-negativity and a sum-to-one constraint, the meta-learner produces interpretable convex weights and mitigates collinearity among base models. A pragmatically tuned clipping policy is introduced to stabilize the conversion from log-space pre- dictions back to prices. In a 10-fold out-of-fold evaluation, the ensemble achieves competitive accuracy (R2 =0.886 in price space; R2 =0.906 in log space), closely tracking the best single model while remaining fully transparent. Beyond accuracy, the method offers operational simplicity: standard, stable hyperparameter ranges; minimal preprocessing; and negligible inference overhead for blending. Taken together, these properties yield a reproducible and reliable blueprint for tabular regression ensembles that balances performance, robustness, and interpretability, and can be adopted with minimal engineering effort in commercial settings.