OptimusPrice

Context

Hotels typically rely on manual pricing strategies or third-party OTAs (Online Travel Agencies) that charge commissions of 15-30%. Manual pricing is slow, reactive, and cannot process the complex demand signals that determine optimal room rates. OTA-dependent pricing surrenders control—and margin—to intermediaries.

The business problem: a mid-sized hotel chain with 50+ properties needs a dynamic pricing engine that maximizes revenue per available room (RevPAR) without requiring a dedicated data science team to operate.

Challenges

• Data sparsity: Historical booking data covers only 18 months with significant seasonal gaps
• Feature engineering: Demand signals must be extracted from booking patterns, competitor pricing, seasonality, and local events
• Model interpretability: Hotel managers need to understand and trust the pricing recommendations
• Real-time inference: Price recommendations must be available in under 2 seconds for front-desk use
• Commission dependency: Reducing OTA reliance without losing bookings requires precision in pricing

Solution

Architecture

Booking History → Feature Engineering → Random Forest → Price API → Streamlit
Competitor     → (Pandas, 40+ feats)  → (scikit-learn)             → Manager UI
Event Calendar                          ↑ Optuna HPO

Pipeline

1. Feature Extraction: Historical booking data is processed through a Pandas pipeline that generates 40+ features including day-of-week, seasonal multipliers, local event proximity, competitor price differentials, and historical occupancy patterns.
2. Model Training: A Random Forest regressor predicts optimal room price. Hyperparameters are tuned using Optuna with 200 trials, optimizing for mean absolute percentage error (MAPE).
3. Inference: The trained model runs as a FastAPI endpoint. Input features are computed in real-time, and the model returns a recommended price plus confidence interval.
4. Dashboard: Hotel managers interact with a Streamlit dashboard that shows current pricing recommendations, historical trends, and model confidence.

Tech Stack

• Python with Pandas for feature engineering and data transformation
• scikit-learn Random Forest regressor for price prediction
• Optuna for automated hyperparameter optimization
• Streamlit for the management dashboard
• Docker for production deployment with CI/CD

Implementation

The critical design choice was opting for Random Forest over deep learning. While a neural network could theoretically capture more complex patterns, the Random Forest provides feature importance scores that hotel managers can interpret. When the model says “price should increase by 15% on Fridays in August,” the manager can see that the decision is driven by historical occupancy (+40% importance) and seasonal factor (+25% importance).

Optuna’s Bayesian optimization was essential. Manual hyperparameter tuning produced models with MAPE around 22%. Optuna reduced this to 14.5% by finding optimal tree depth, minimum samples per leaf, and feature sampling strategies.

The Streamlit dashboard includes a confidence indicator. When the model is uncertain (wide prediction interval), the dashboard shows a range rather than a single number. This prevents over-reliance on the model and keeps the human in the loop.

Trade-offs

Random Forest was chosen over gradient boosting (XGBoost, LightGBM) despite the latter typically achieving higher accuracy on tabular data. The deciding factor was interpretability: feature importance scores from Random Forest are more stable and easier to explain to non-technical stakeholders. In a business context where hotel managers must trust and act on model recommendations, interpretability outweighed a marginal accuracy gain.

The model uses historical booking data without real-time demand signals (e.g., current booking pace, competitor live rates). Incorporating streaming data would improve accuracy during high-demand events but requires integration with OTA APIs and property management systems — a significant engineering investment that makes sense once the batch model has proven value.

Reliability

The model includes automated retraining triggers. When mean absolute percentage error exceeds 20% on recent predictions, the pipeline triggers a new training cycle with updated data. This compensates for market shifts without manual intervention.

Input validation rejects out-of-range features before they reach the model. If occupancy rate exceeds 100% or price falls below a configured floor, the system logs the anomaly and returns the last known recommendation instead of a potentially incorrect prediction.

Results

• R² = 0.92 on price prediction, with MAPE of 14.5%. Based on historical test data with 200 Optuna hyperparameter trials.
• Backtesting simulation projected a net revenue improvement of approximately 30% over the OTA-baseline scenario (30% commission). This is a modeled estimate, not a measured A/B result — actual performance depends on market conditions and adoption.
• Inference latency under 500ms including feature computation, meeting the real-time requirement for front-desk use.
• Direct booking capability: the model enables competitive direct pricing against OTA rates while retaining full margin — the primary business value.

Lessons Learned

1. Interpretability is a feature, not a limitation. The Random Forest’s feature importance scores were cited by hotel managers as the reason they trusted the system.
2. Hyperparameter optimization should be automated. Manual tuning cost weeks of iteration. Optuna found a better model in 4 hours.
3. Confidence intervals matter more than point predictions. Hotel managers preferred knowing when the model was uncertain over getting a single number that might be wrong.

Future Work

• Incorporate real-time events (local festivals, conferences) via API integration
• Multi-property model that shares learnings across the chain
• A/B testing framework to measure actual revenue impact against manual pricing