Machine Learning-Based Prediction of Hotel Guest Ratings: Price Associations and Geographical Transferability in Cairo, Egypt
DOI:
https://doi.org/10.5281/zenodo.24389874Keywords:
accommodation ratings; Cairo; Booking.com; quoted prices; geographical transfer; machine learning; data provenance; service attributes.Abstract
The accommodation scores that can be found online and the room prices that are quoted signal perceived quality, but the predictive relationship differs depending on market composition and validation setting. This study investigates a file made available by Booking.com of 500 accommodation listings in Cairo with a stay search for 1–20 February 2025. The resulting dataset contained 477 rated listings which we retained. The quoted context pertains to a date range of 2022–2026, which fits within a 5-calendar-year search; the review dates and extraction timestamp are missing. Using five repeated five-fold splits at the property level, and a two-way Giza-versus-other-Cairo transfer test, we compare a median baseline, ridge regression, random forest and extra trees. The HSCO residuals have been used in the analysis. Random forest attained a MAE of 0.622 rating points (1-10 scale). The baseline value was 0.732, so we can reduce 15.1%. Under geographical transfer, its MAE increases to 0.878 and its R² becomes negative. The correlation between price and rating is weak on the basis of the unadjusted coefficient (Spearman ρ = −0.061). The adjusted coefficient per doubling of quoted price is 0.161 points (95% interval 0.074–0.249). After duplicate removal, a separately analysed and undated Egypt Kaggle corpus provides 561 rated property proxies, whose scores and service counts are not pooled with Cairo. The findings highlight slight within-snapshot predictability and considerable transfer sensitivity, as opposed to a causal price premium, sentiment-classification advance, or five-year trend. Units, rating scales and temporal evidence are retained throughout