A Predictive and Interpretable Machine Learning Approach for Revenue-Optimized Airbnb Pricing

"This thesis develops a data-driven pricing framework for Airbnb listings in Amsterdam that predicts booking probability, optimizes nightly prices to maximize expected revenue, and explains each recommendation using interpretable machine learning. The framework is built on publicly available Inside Airbnb data covering 6,993 listings and 2.55 million listing-day observations over one year. The core of the approach is a monotonically constrained LightGBM model trained to predict whether a listing will be booked on a given date at a given price. Monotonic constraints on two price-related features ensure the model respects the economic principle that higher prices should not increase demand. The model’s predicted probabilities are calibrated using isotonic regression and then used inside an expected revenue formula, where a constrained optimizer sweeps candidate prices within ±15% of each listing’s historical price and selects the revenue-maximizing option. The final model achieves a test AUC of 0.755 and a calibrated Expected Calibration Error of 0.102. At the ±15% constraint level, the model-optimized strategy produces a median expected revenue of €183.3 per listing-night, compared to €166.9 for the host’s historical price — an uplift of 9.8%. Of the 6,993 listings, 74% receive a price increase recommendation, 12% a decrease, and 14% are advised to hold. An independent review-based sanity check using data the model never saw confirms that the recommendations are directionally consistent with guest perceptions across all seven Airbnb review categories. SHAP-based explanations provide transparency at both the global and individual listing level, and a static host-facing prototype demonstrates how these explanations can be presented in plain language. Although the prediction model operates at the listing-day level, the optimization results in this thesis are reported as one representative recommendation per listing under typical test-period conditions. The framework is positioned as constrained decision support rather than a causal pricing engine, reflecting the inherent limitations of observational data — particularly proxy label noise and residual price endogeneity. The methodology is fully reproducible using open data and open-source tools".