Real-estate valuation depends on more than isolated property characteristics. Price emerges from nonlinear interactions among living space, construction quality, location, age, renovation history, waterfront access, and neighborhood context.
This project evaluates whether a feedforward neural network can model those interactions more effectively than interpretable linear and ridge-regression baselines. All models are trained and evaluated through a single validation pipeline that keeps the target on a common scale, fits preprocessing inside each training fold to prevent leakage, and reports performance on both a held-out test set and five-fold cross-validation.
THE CHALLENGE
Housing prices are shaped by relationships that are rarely purely linear. Additional square footage may have different value depending on construction grade, neighborhood, geographic location, renovation status, and surrounding property characteristics.
The central challenge was not simply to train a more complex model. It was to determine whether that additional complexity produced stable, repeatable improvements on unseen properties while maintaining a fair comparison with simpler baselines.
PROJECT AT A GLANCE
Methodology
&
Insights
THE APPROACH
I created a reproducible 70/30 train-test split and reserved the test set exclusively for final evaluation. Sale dates were converted into calendar features, and additional variables captured property age, renovation status, and years since renovation. ZIP code was treated as a categorical variable, while latitude and longitude preserved continuous geographic information.
All preprocessing was learned from the training data only. Numeric features were standardized, categorical features were one-hot encoded, and the target was standardized during neural-network training before predictions were converted back into U.S. dollars.
Neural-network architectures were compared using a separate validation subset. The selected feedforward network used two hidden layers, ReLU activation, L2 regularization, mini-batch optimization, and early stopping. Linear and ridge regression served as transparent benchmarks.
RESULTS
On the untouched test set, the neural network achieved an R² of 0.881, an RMSE of $123,657, an MAE of $71,262, and a mean absolute percentage error of 13.6%.
Compared with linear regression, the neural network reduced test RMSE by 19.4% and MAE by 24.4%. The result was also consistent across five-fold cross-validation, where the network achieved an average RMSE of $121,256 ± $4,231 and an average R² of 0.891. Its cross-validated RMSE was 24.7% lower than the linear baseline.
These results provide evidence that nonlinear modeling captured meaningful interactions among property, neighborhood, and geographic features that were not represented as effectively by an additive linear model.
LIMITATIONS AND NEXT STEPS
The dataset represents sales from 2014 and 2015 and should not be interpreted as a model for current home values without retraining on recent market data. It also excludes important external variables such as interest rates, school quality, neighborhood amenities, local inventory, economic conditions, and property-specific renovations.
Performance was weaker for unusually expensive properties, where prediction errors had greater dollar magnitude. The neural network also provides less direct interpretability than linear regression. A production version would require feature-attribution methods, geographic and temporal drift monitoring, subgroup error analysis, prediction intervals, and scheduled retraining as market conditions change.