Forecasting hotel room revenue like a time-traveling data wizard.
From old-school ARIMA to state-of-the-art attention-based deep learners — all in one notebook!
This project is carried out under the guidance of Dr. Arjun Ray.
📈 Prediction Gallery
| Forecast Type | Preview |
|---|---|
| Prophet 1-Year Daily Model | .png) |
| Scatter Plot- 1 day |  |
| TBATS |  |
| Prophet 1-Month Daily Model | .png) |
| Prophet 3-Month Daily Model | .png) |
✅ Tried-and-tested classical models
✅ ML regressors + exogenous variable magic
✅ Neural networks with custom attention
✅ 1-day, 30-day, and 365-day forecasts
✅ Easy to reproduce + exportable plots
✅ 📈 Prophet with daily learning turns out to be the best performer for long-term forecasts (1 year)!
When trained every day and set for 1-year windows, Prophet dominates the leaderboard!
Thanks to its intuitive handling of seasonality, holidays, and trends — it's ideal for hotel data.
📊 Visual: Prophet Forecast vs Reality
- 🔢 ARIMA/SARIMA
- 🤖 Auto-ARIMA + Exogenous SARIMAX
- 📈 Ridge Regression / XGBoost
- 🧬 LSTM / GRU with Attention
- 🧪 Ensembles
- ⏳ TBATS / Decomposition
- 🧠 VAR (Multivariate)
- 🧿 Prophet (Daily retrained)
- 🧠 BiLSTM + GRU + Custom Attention
📦 DeepForecast-Revenue
├── model_train_classification_4_files.ipynb
├── data/
│ ├── df_4_files_combined_no_outliers.pkl
│ └── df_4_files_combined_no_outliers_for_AR.pkl
├── images/
│ ├── prophet_1_year_forecast.png
│ ├── bilstm_attention_forecast.png
│ └── ...
├── models/
│ ├── best_arima_model.pkl
│ ├── prophet_model_365.pkl
├── outputs/
│ ├── 30_day_forecast_ARIMA.png
│ └── ...
conda create -n forecast_env python=3.8
conda activate forecast_env
pip install -r requirements.txtOr install manually:
pip install numpy pandas matplotlib seaborn statsmodels pmdarima scikit-learn xgboost prophet tensorflow pytorch-lightning pytorch-forecasting tbats holidays joblibjupyter notebook model_train_classification_4_files.ipynb🛫 Start with: Clean daily hotel room revenue 🔄 Transformations: Resampling, decomposition 📊 Forecasting methods:
- ARIMA → SARIMA → Auto-ARIMA
- SARIMAX + Exogenous features (ARR, Rooms Sold, etc.)
- Ridge & XGBoost
- Prophet with Holidays
- LSTM / GRU / BiLSTM + Attention
- TBATS + VAR
- 🧠 Ensemble learning for robust results!
📦 Outputs Saved:
- Forecast PNGs
- Residual plots
- Joblib models
- Cross-validation MAE/RMSE
Prophet, when retrained every day and tuned for a 1-year forecast window, delivers the lowest MAPE and most stable predictions — especially during seasonal spikes and sudden drops.
| Model | Horizon | MAPE ↓ | RMSE ↓ |
|---|---|---|---|
| Prophet (Daily) | 365 | 8.7% | 1243 |
| XGBoost | 30 | 10.5% | 1460 |
| SARIMAX + Exog | 90 | 11.9% | 1522 |
| Ensemble (All) | 30 | 9.2% | 1330 |
📌 Visual comparison plots saved in images/.
This project was born out of a desire to build an all-in-one forecasting Swiss army knife for hotels. Whether it’s a budget inn or a five-star chain — daily revenue prediction matters.
🧙♂️ Predict the future... one revenue spike at a time. — Team DeepForecast