Demand forecasting is the foundation of revenue management: if you don't know how many guests are coming, every subsequent decision about pricing, staffing, inventory, and F&B planning is built on uncertainty. This paper provides a rigorous systematic review of how AI-based forecasting has changed the game — evaluating which algorithms work best, under what conditions, and what the barriers to adoption look like in practice.
The paper reviews a substantial body of literature and distills the performance evidence clearly. Across nearly all market conditions and data configurations tested, AI-based forecasting models outperform traditional statistical methods — primarily Exponential Smoothing (used in most legacy revenue management systems) and ARIMA models. The performance gap is most significant in three specific scenarios: markets with high demand volatility (where traditional models struggle to adjust quickly to pattern changes), demand influenced by irregular events (large conferences, local festivals, unusual weather), and properties with access to rich data inputs beyond internal booking history (competitor rates, search data, economic indicators).
The taxonomy of AI methods the paper develops is genuinely useful for non-technical readers. Machine learning methods — gradient boosting, random forest, support vector regression — tend to perform best when feature engineering is done carefully (i.e., when the right input variables are selected and prepared by someone who understands what drives demand in that market). Deep learning methods — LSTM networks in particular — perform best with longer time horizons and large amounts of historical data, making them most suitable for chains or managed properties with multi-year clean booking histories. Hybrid approaches combining elements of both often outperform either alone.
The paper also documents what the best-performing implementations have in common, beyond algorithm choice. Data preparation is repeatedly identified as the most critical factor: models trained on messy, incomplete, or poorly formatted data will underperform simpler models trained on clean data. The investment in data infrastructure — consistent PMS configurations, clean historical rate and booking records, reliable external data feeds — pays dividends that exceed the investment in the model itself.
For operators making technology decisions, the paper offers a useful reality check on vendor claims. "AI-powered forecasting" has become a ubiquitous marketing phrase. The paper provides the technical vocabulary and performance benchmarks to ask meaningful questions: What algorithm is the model based on? What training data was it developed on? What was the accuracy improvement over your existing baseline? On what properties, in what markets, over what time period was that performance measured?
The practical advice is measured and realistic. Not every hotel needs a custom deep learning model. For many independent properties and smaller chains, well-implemented commercial forecasting tools with ML components will deliver meaningful performance improvements at reasonable cost. The larger opportunity — and the larger investment requirement — lies in hotels with the data maturity and analytical talent to customize and continuously improve their models. The paper is clear that this is a journey, not a one-time deployment: AI forecasting systems improve over time with more data, more feedback cycles, and more organizational learning about how to interpret and act on their outputs.