Income Prediction Neural Network
Oct 2025
View on GitHubWhat I Did
Built a 2-layer neural network to predict income levels from US Census data. The goal was to systematically test what actually improved performance versus what just looked sophisticated on paper.
Approach
Instead of using one model, I ensembled 5 cross-validation models to reduce variance. Optimized decision thresholds to handle class imbalance (most people earn under $50K). Stripped out features that added noise rather than signal.
Results
Achieved 85.6% accuracy. But honestly, the process taught me more than the final number. Removing complexity often improved performance. Adding more features didn't always help. Sometimes the best optimization is figuring out what not to do.
What I Learned
Complexity doesn't equal performance. Half the work was removing things that seemed like they should help but actually hurt the model. That's harder than it sounds because it goes against the instinct to keep adding more.