Hotel Rating Prediction from Online Guest Feedback Using Reliability Modeling and Neural Text Analysis

Online hotel reviews provide a major source of information for understanding guest satisfaction, yet rating prediction remains difficult because review text is often short, highly skewed toward positive scores, and affected by inconsistent or repetitive content. This study presents a framework designed to prevent information leakage in hotel rating prediction. The empirical analysis uses 26,675 Booking.com reviews covering 819 hotels. After minimal cleaning, 26,384 reviews are retained and split chronologically before any vectorizer fitting, reliability scoring, or model training. The protocol ensures that the test set is never filtered and that no rating-derived reliability signal is used to construct training features. The study evaluates TF-IDF Ridge baselines, recurrent neural models, and transformer architectures including DistilBERT, BERT, RoBERTa, MobileBERT, and TinyBERT. Evaluation is performed on the same unfiltered chronological test set using MAE, RMSE, R2, rounded rating accuracy, accuracy within one rating point, low score recall, rating group errors, and confusion matrices. DistilBERT achieves the strongest overall performance, with MAE = 0.4370, RMSE = 0.6979, and R2=0.8217, while BiLSTM models show stronger sensitivity to low rating reviews. Additional analyses include language composition auditing, unseen hotel generalization, reliability threshold sensitivity, and forecasting of future hotel rating dynamics across 7-, 30-, 60-, and 90-day horizons. The results show that reliable hotel rating prediction requires both expressive language models and careful evaluation protocols that separate review reliability analysis from filtering based on the target rating.