Portugal Inflation Forecast: A Practical Guide to Time Series Modeling Using Machine Learning

• Original Author/Maintainer: Viksom
• Source Platform: GitHub
• Original Project Name: Inflation-Forecast-Using-Machine-Learning-Models
• Original Link: https://github.com/Viksom/Inflation-Forecast-Using-Machine-Learning-Models
• Publication Date: May 2025

This article focuses on Portugal's inflation rate prediction, integrating multi-source economic data, combining time series analysis with models like ARIMA, XGBoost, and LightGBM to build a complete prediction workflow, providing engineering practice references for macroeconomic time series forecasting.

Inflation is a core macroeconomic indicator that affects monetary policy, investment decisions, and residents' living costs. Traditional econometric models (such as ARIMA, VAR) have limitations in handling nonlinear and high-dimensional features, while machine learning techniques can capture complex patterns to improve prediction accuracy. This project demonstrates the technical practice of Portugal's inflation forecast through a complete pipeline.

Multi-source Data System

Price and Cost Indicators: CPI and its sub-items, PPI, import price index, energy and commodity prices Economic Activity Indicators: GDP and its growth rate, industrial value-added, retail sales, unemployment rate Monetary and Financial Indicators: M1/M2, interest rate levels, exchange rate fluctuations External Factors: EU economic sentiment index, global supply chain pressure index, geopolitical risk index

Integration Process

• Data cleaning: Missing value interpolation/forward filling, outlier correction, unified monthly frequency
• Feature alignment: Time dimension consistency processing
• Lag features: Construct multi-period lag values to capture historical impacts

Project Structure

12 ordered Jupyter Notebooks covering the entire workflow of data preparation, integration, analysis, feature engineering, model training, and evaluation.

Core Algorithms

• Statistical benchmarks: ARIMA/SARIMA, exponential smoothing, VAR
• Machine learning: Random Forest, XGBoost/LightGBM, SVR
• Deep learning: LSTM/GRU, Transformer

Hyperparameter Optimization

Bayesian optimization using the Optuna framework to efficiently explore optimal parameter combinations.

Stationarity Processing

• Differencing transformation: Eliminate trends and seasonality
• Log transformation: Stabilize variance and convert growth rates
• Seasonal decomposition: X-13ARIMA-SEATS or STL decomposition for trend, seasonality, and residuals

Feature Engineering

• Lag features: Capture inflation inertia
• Rolling statistics: Sliding window mean/standard deviation, etc.
• Time features: Month/quarter/year
• Interaction features: Synergistic effects between economic indicators

Evaluation Metrics

• Error metrics: RMSE (sensitive to large errors), MAE (robust), MAPE (cross-series comparison)
• Directional accuracy: Correct rate of up/down prediction, Theil's U (advantage over naive prediction)

Cross-Validation

• Rolling window validation: Forward chaining to expand the training set
• TimeSeriesSplit: Ensure training set is earlier than validation set to avoid data leakage

Challenges

• Structural changes (Lucas critique), data limitations (low frequency/revisions), multi-factor interweaving, policy interventions

Value and Limitations of ML

• Value: Nonlinear capture, automatic feature selection, ensemble learning, high-dimensional processing
• Limitations: Data dependence, black-box nature, overfitting risk, lack of causal explanation

Recommendations

• Model combination strategy, uncertainty quantification, real-time update mechanism, domain knowledge integration

This project has built a complete inflation prediction pipeline, and its technical architecture and practices have reference value for similar tasks. In the future, we can explore high-frequency data, Nowcasting models, interpretation tools, and real-time prediction systems. Macroeconomic forecasting needs to combine ML tools with economic logic to find patterns in uncertainty.