Machine Learning-Based Forecasting of Headline Inflation in Ethiopia: an Empirical Comparative Study

Technical Summary: Machine Learning-Based Forecasting of Headline Inflation in Ethiopia

Problem Statement
Inflation remains a critical determinant of national welfare and a significant economic challenge in Ethiopia, where headline inflation has recently exceeded 30%, driven by factors such as drought, internal conflict, and global supply shocks. Traditional econometric and time series models (e.g., ARIMA, VAR, VECM) have historically been used to forecast inflation in Ethiopia. However, these linear models often struggle with accuracy due to the complex, nonlinear nature of inflation data and their inability to adequately capture structural breaks, asymmetries, and the multifaceted interactions of macroeconomic variables. Furthermore, previous studies in the region have frequently relied on aggregate Consumer Price Index (CPI) data without sufficiently accounting for specific determinants or utilizing advanced machine learning (ML) techniques to improve forecast precision.

Methodology
This study employs a supervised machine learning approach to model and forecast headline inflation in Ethiopia using quarterly data spanning from 2000 to 2023. Data was sourced from the Central Statistics Service, National Bank of Ethiopia, Ethiopian National Meteorological Agency, and the World Bank.

• Data Preprocessing: The dataset was cleaned, transformed, and split into an 80% training set and a 20% test set. The response variable, Headline Inflation ($HI_t$), was calculated as the natural logarithm of the year-over-year percentage change in the Consumer Price Index.
• Predictor Variables: The study incorporated a wide range of exogenous variables, including food and non-food inflation, money supply, exchange rates, GDP, government budget deficits, foreign direct investment, rainfall, oil and food prices, political stability, transport data (number of vehicles), and unemployment rates.
• Modeling Approaches: The study compared several supervised learning algorithms against traditional benchmarks:
  • Penalized Linear Models: Ridge Regression, LASSO (Least Absolute Shrinkage and Selection Operator), and Elastic Net to address multicollinearity and perform variable selection.
  • Nonlinear Machine Learning Models: Random Forest, Multi-Layer Perceptron (MLP), and Neural Network Autoregression (NNAR).
  • Benchmark: Univariate SARIMA models were used as the baseline for comparison.
• Evaluation Metrics: Model performance was assessed using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) for both in-sample (training) and out-of-sample (testing) data.
• Forecasting Horizon: The selected optimal model was used to forecast headline inflation for a five-year horizon (2024–2028).

Key Results

• Feature Importance: Across all models, food inflation was identified as the most critical predictor. Other top variables included non-food inflation, export prices, government consumption, gross fixed investment, and the number of vehicles. Notably, the LASSO model identified "political stability" and "number of vehicles" as significant factors, with the latter showing a negative correlation with inflation, suggesting improved transport efficiency may alleviate price pressures.
• Model Performance:
  • Linear Models: LASSO regression effectively addressed multicollinearity, achieving strong results with an in-sample RMSE of 0.110 and an out-of-sample RMSE of 0.172. It outperformed Ridge and Elastic Net models in out-of-sample forecasting.
  • Nonlinear Models: The Neural Network Autoregression (NNAR) model, specifically the NNAR(5, 2, 10) with exogenous variables, outperformed all other models. It achieved perfect in-sample accuracy (RMSE = 0.000) and the lowest out-of-sample error (RMSE = 0.128; MAE = 0.067; MAPE = 10.576).
  • Comparison: The NNAR model demonstrated superior generalization capabilities compared to SARIMA, MLP, Random Forest, and linear shrinkage models. The inclusion of exogenous variables significantly improved forecast accuracy compared to univariate approaches.
• Forecasting Outcome: The NNAR model predicted that headline inflation would remain relatively low in 2024 and 2025, gradually increase starting in 2026, peak at approximately 36% in the second quarter of 2027, and then begin a gradual decline by the fourth quarter of 2028.
• Diagnostics: Residual analysis (histograms and ACF plots) and Ljung-Box tests confirmed that the NNAR model residuals were normally distributed with no significant serial correlation, indicating the model adequately captured the underlying data structure.

Significance and Contributions
The study claims to be the first in Ethiopia to utilize machine learning approaches specifically for forecasting headline inflation. Its primary contributions include:

1. Methodological Superiority: Demonstrating that nonlinear machine learning models, particularly NNAR, offer superior forecasting performance over traditional linear econometric models and univariate time series methods in the context of Ethiopia's volatile inflation environment.
2. Variable Identification: Validating the importance of diverse predictors, including political stability and transport infrastructure (number of vehicles), which are often overlooked in traditional inflation modeling. The identification of vehicle access as a factor reducing inflationary pressure is presented as a novel finding.
3. Policy Relevance: By providing more accurate forecasts and identifying key drivers, the study offers better tools for economic planning and monetary policy decision-making, aiding the National Bank of Ethiopia in its goal of maintaining price stability.
4. Robustness: The study confirms that incorporating exogenous variables into ML models significantly enhances forecast accuracy, addressing the limitations of models that rely solely on historical inflation data.