House price prediction from historical sales, end to end

You have historical home sales with features such as lot area, year built, number of rooms, neighbourhood and sale date. Design an end-to-end approach that predicts the sale price of new listings, from framing the problem to a lightweight model in production.

Requirements

• Functional: a price prediction for each new listing; an explanation of what drives it.
• Non-functional: accurate across price ranges and neighbourhoods; robust to market shifts over time; fair across areas; small and fast enough for the serving budget.

Clarifying questions

• How many sales, over what period and in which regions?
• What is the prediction used for: pricing advice, risk, search ranking?
• Which error matters more: absolute dollars or percentage?
• How fast does the market move, and how often can we retrain?
• What latency and memory budget at serving time?

What a strong answer covers

• Framing as regression on log(price), so errors are relative and the target is less skewed.
• Cleaning: missing values imputed with care (and a missing indicator), outliers checked rather than dropped blindly, and leakage avoided (no features known only after sale, such as the final assessed value).
• Features: age at sale, size per room, neighbourhood encodings (target encoding fitted inside folds), location (latitude and longitude, distance to centres), sale month and market index, useful interactions.
• Models: a regularized linear baseline, then gradient-boosted trees; ensembles only if they earn their cost.
• Evaluation: time-based splits (train on the past, test on later sales), RMSE on log price and MAPE, per-segment error; hyperparameter tuning inside the time-based folds and error analysis on the worst segments.
• Interpretability (SHAP values, partial dependence) and fairness checks across neighbourhoods.
• Non-stationarity: monitoring error and feature drift, retraining on a schedule, and adjusting for market index changes.
• Deployment: a compact tree model or linear model with a small feature pipeline under the latency and memory limits.

Follow-ups

• Prices fall 10% in three months. What breaks first, and how do you notice?
• Predict a price range, not a point.
• A new city with very little history.
• Explain one specific prediction to a home seller.