Hotel Price Prediction Based on Multimodal Learning: Applications of MLP, LSTM, and Fusion Models

Original Author/Maintainer: KMS92-L, Source Platform: GitHub, Project Title: hotel-price-prediction-multimodal. The core of this project is to use multimodal learning methods combined with MLP, LSTM, and fusion models, integrating macroeconomic indicators (such as CPI, crude oil prices) and historical hotel price data to achieve more accurate hotel price prediction and provide data-driven decision support for revenue management.

Hotel price prediction is a core issue in revenue management, affecting hotel profitability and consumer decisions. Traditional methods rely on single data sources or simple time series models, making it difficult to capture complex factors. Multimodal learning builds comprehensive models by integrating data from different sources. This project explores combining macroeconomic indicators with historical price data to improve prediction accuracy using deep learning.

Macroeconomic Indicators: Integrate CPIAUCSL.csv (Consumer Price Index), CUSR0000SEHF01.csv (CPI sub-item for hotel accommodation), and MCOILWTICO.csv (crude oil price index) to provide macro-environment context. Hotel Price Data: Serves as the target variable and basic feature.

• MLP: Baseline model that captures non-linear relationships between features and learns the mapping from inputs to prices via a fully connected network. - LSTM: Handles time series characteristics; its gating mechanism captures long-term trends and seasonality, avoiding gradient vanishing. - Fusion Model: Core innovation, which may adopt early (input layer concatenation), mid-term (post-encoding fusion), or late (weighted after independent prediction) fusion strategies to capture both macro trends and micro fluctuations.

Data Preprocessing: Time series alignment, missing value interpolation/forward filling, indicator standardization. Feature Engineering: Construct price lag terms, rolling statistics (moving average/standard deviation), macroeconomic derived features (inflation rate, price change rate). Model Training: Time series cross-validation (to prevent data leakage), early stopping mechanism (to prevent overfitting), hyperparameter tuning (network structure, learning rate, etc.).

• Dynamic Pricing: Predict price trends, formulate peak price increases and off-season discount strategies to maximize revenue. - Inventory Management: Optimize room allocation by combining demand forecasting, balancing supply between direct sales and OTA channels. - Investment Decisions: Assist hotel investors in evaluating potential returns of regional/type hotels. - Consumer Assistance: Provide booking timing recommendations to help users get discounted prices.

• Multimodal Versatility: Can be extended to fields such as airfare, housing price, and stock price prediction. - Combination of Deep Learning and Traditional Methods: Choose models based on data characteristics instead of blindly pursuing complex architectures. - Value of External Data: Macroeconomic data significantly improves prediction capabilities; relevant external signals should be introduced when building models.

This project demonstrates the potential of multimodal learning in hotel price prediction by integrating MLP, LSTM, and fusion models. By combining macroeconomic and historical price data to capture multiple influencing factors, it provides decision support for revenue management. It offers a valuable reference implementation for developers/researchers in time series prediction, revenue management, or multimodal learning.