XGBoost-Based Time Series Forecasting with Bootstrapped Confidence Intervals

Overview

This script performs time series forecasting for revenue data using XGBoost regression, incorporating rolling forecast validation, hyperparameter tuning, and bootstrapped confidence intervals. It forecasts the next 12 months of revenue and visualizes the predictions with uncertainty bounds.

The key file for running this script is main.ipynb.

---

1. Load and Prepare Data

• Reads an Excel file (Connect_analyst_priklad.xlsx)
• Converts the ACCOUNTING_MONTH_KEY column into a datetime format.
• Aggregates revenue at a monthly level.
• Creates time-based features:
  • Year
  • Month
  • Year_squared (to capture potential quadratic trends in revenue).

---

2. Define Features and Target

• X (Features): Year, Month, Year_squared
• y (Target): Revenue

---

3. Rolling Forecast Split

A custom function (rolling_forecast_split) is implemented to:

• Perform expanding window validation, where the model is trained on increasing historical data.
• Uses an initial training window of 12 months, expanding by 3 months at each step.

---

4. Hyperparameter Tuning for XGBoost

• Uses GridSearchCV to find the best parameters for the XGBRegressor model.
• Parameters tuned:
  • n_estimators: [300, 500]
  • learning_rate: [0.01, 0.05]
  • max_depth: [3, 5]
• The best model is selected based on Mean Absolute Error (MAE).

---

5. Model Evaluation

• The rolling forecast approach is used to evaluate model performance.
• Computes multiple error metrics:
  • Mean Absolute Error (MAE)
  • Root Mean Squared Error (RMSE)
  • Mean Absolute Percentage Error (MAPE)
  • Mean Percentage Error (MPE)
  • R-squared (R²)

---

6. Forecasting the Next 12 Months

• Generates future dates for the next 12 months.
• Uses the trained XGBoost model to predict revenue for future periods.

---

7. Bootstrapped Confidence Intervals

• Bootstrap sampling (1000 iterations) is used to estimate the uncertainty of predictions.
• Generates upper and lower bounds (95% confidence intervals).

---

8. Visualization of Forecast

• Plots historical revenue alongside forecasted values.
• Includes confidence intervals to quantify prediction uncertainty.

---

Final Output

• A DataFrame displaying predicted revenue and confidence intervals.
• A time series plot comparing historical data with forecasted values.

---