Context
The client has forecasts
Challenge
Predict average intervention price by product group (120 predictions), by month and over 18 months
Technologies used
DBMS: MS SQL Server 2016 • Python 3.6 • Frameworks: Pandas, Scikit-Learn, FBProphet, XGBoost
Methodology and implementation
Machine learning project for intervention cost forecasting:
**Approach** 1. Exploration of historical data (3 years of data) 2. Feature engineering: extraction of temporal and categorical variables 3. Testing of several algorithms: ARIMA, FBProphet, XGBoost, LSTM 4. Temporal cross-validation (time series split) 5. Hyperparameter optimization 6. Performance evaluation (MAPE, RMSE)
**FBProphet Methodology** Using FBProphet to capture: - Long-term trends - Annual and monthly seasonality - Holidays and special events - Change point detection
**Production deployment** - Model retrained monthly with new data - Automated pipeline in SQL Server - Dashboard for monitoring predictions vs actual
Results
Between 4 and 7% error on 18-month prediction (cross-validation), for the 3 main product categories. Model in production, retrained every month
Visualizations
Average intervention cost forecasts for two product classes
Decomposition of components in FBProphet (trend, seasonality)