Our Pricing Algorithm: Methodology & Sources

How Demand Forecasting Works

Nexorev's demand forecasting pipeline uses an ensemble approach that combines multiple statistical and machine learning models rather than relying on a single algorithm. The primary models are Prophet (Meta's additive time-series forecasting framework), ARIMA (autoregressive integrated moving-average models via statsmodels), and XGBoost (gradient-boosted decision trees). Each model captures different aspects of hotel demand: Prophet handles trend and seasonality decomposition, ARIMA captures short-term autocorrelation patterns, and XGBoost integrates tabular features like day-of-week, event proximity, and competitor rate positioning.

The ensemble combines predictions from all three models using a weighted average, where weights are dynamically adjusted based on each model's recent accuracy on held-out validation data. This approach is more robust than any single model because it reduces the risk of systematic errors. If one model overestimates demand during a particular pattern such as a holiday weekend, the others can correct the aggregate prediction. The specific weighting methodology and validation protocol are documented in the model performance page.

Data Sources and Their Limitations

The current backtest data comes from public sources: ISTAT (Italian national statistics office) provides monthly tourism arrival and occupancy data by region. Banca d'Italia publishes international tourism expenditure surveys. ENIT (Italian National Tourism Agency) releases seasonal demand reports and regional tourism indicators. STR (Smith Travel Research) provides industry methodology references and benchmark definitions. These sources inform the market priors and seasonal patterns that the forecasting models learn from.

Public data has inherent limitations. ISTAT publishes at monthly granularity with a reporting delay of several months, meaning the forecasting models cannot rely on it for real-time demand sensing. Hotel-level data such as actual bookings, cancellations, room-type mix, and guest segments is only available through PMS integration with individual properties. The transition from public-data backtests to PMS-fed live forecasting is the core objective of the pilot program. Until that transition is complete, all forecast accuracy metrics should be interpreted as public-data performance, not production hotel performance.

How Dynamic Pricing Recommendations Are Generated

The pricing recommendation engine takes demand forecasts as input and generates suggested nightly rates for each room type and rate plan. The engine operates within a constraint framework defined by the hotel: minimum rates (floors), maximum rates (ceilings), blackout periods, event-specific rules, and human approval thresholds. No rate recommendation is published to distribution channels without passing through these guardrails.

The pricing logic considers multiple factors simultaneously: forecasted occupancy for the target date, current booking pace versus historical pace for the same lead time, competitor rate positioning within the defined competitive set, day-of-week demand patterns, local event calendars, and length-of-stay optimization targets. The output is not a single optimal price but a recommended rate range with a confidence score, allowing the hotel operator to make the final decision with full visibility into the reasoning.

In pilot mode, all recommendations are delivered through a dashboard interface for human review and approval. Autonomous rate updates, where the system pushes rates directly to the channel manager, are available but should only be activated after the hotel has validated recommendation quality over a meaningful test period.

Adapting the Models for US Hotel Markets

For US hotel markets, the forecasting pipeline adapts to different data sources and demand patterns. US-specific inputs include federal and state holiday calendars, major event and convention schedules, and publicly available occupancy trend data from STR and AHLA (American Hotel and Lodging Association). The models also account for US-specific patterns such as spring break seasonality, summer road-trip demand curves, and business travel cycles that differ significantly from European markets.

PMS integration for US properties follows the same methodology: the system starts with publicly available market-level data to establish baseline demand patterns, then refines forecasts progressively as property-specific PMS data becomes available. This staged approach allows a US hotel to begin receiving useful pricing recommendations within days of starting a pilot, with accuracy improving as the model trains on the property's actual booking data.

Validation and Backtest Protocol

All backtest results published on this site follow a walk-forward validation protocol. The training data is split chronologically: models are trained on historical periods and tested on subsequent periods that the model has never seen. This prevents data leakage and provides a realistic estimate of how the model would have performed if deployed in real time.

Backtest metrics include Mean Absolute Percentage Error (MAPE) for occupancy forecasts, Root Mean Squared Error (RMSE) in occupancy percentage points, and simulated RevPAR lift calculated by applying dynamic pricing rules to historical demand patterns. These metrics are published with their calculation methodology on the model performance page. When production hotels generate real performance data, those results will be published separately with clear labeling distinguishing them from backtest metrics.