修补水晶球：用机器学习增强通胀预测（英）

Mending the Crystal Ball: Enhanced Inflation Forecasts with Machine Learning

Introduction

Forecasting inflation in Japan has become increasingly challenging since 2022. For the first time in three decades, inflation rose above the two-percent target for more than two years, starting from April 2022. Initially driven by cost-push factors, inflation is now increasingly influenced by demand-side factors, particularly in services. The Bank of Japan (BOJ) has repeatedly revised its inflation forecasts upward, reflecting the surprise caused by these developments.

The challenge extends beyond the BOJ, as many central banks were caught off guard by inflation in 2021 and 2022. The World Economic Outlook has consistently underestimated inflation in advanced economies, highlighting the need for improved forecasting methods. The pandemic has introduced structural changes in price dynamics that traditional models failed to capture, leading to unexpected supply chain constraints and a steeper Phillips curve in many industrialized countries.

Methodology

This paper applies machine learning (ML) models to forecast near-term core inflation in Japan, which excludes fresh food and energy prices. The forecasting target includes one-month, three-month, and six-month ahead horizons to maximize the training sample period. Four ML models—two penalized regression models (LASSO and Elastic Net) and two tree-based models (Xgboost and Random Forest)—are compared against two benchmark models.

Potential Predictors of Inflation (Features)

Key predictors identified for post-2022 inflation include:

• Household inflation expectations
• Inbound tourism
• Exchange rates
• The output gap

Model Performance

For the year 2023, the two penalized regression models systematically outperformed the benchmark models, with LASSO providing the most accurate forecast. These models effectively captured the non-linear relationships and emerging drivers of inflation during the pandemic.

Conclusion and Future Work

This study contributes to the growing literature on macroeconomic forecasting using ML models. The use of ML models highlights their ability to incorporate a wide range of variables, handle non-linear relationships, and focus on out-of-sample performance. Future work should continue to refine these models and explore their application to other economic indicators.

Key Findings

• Inflation Forecasting: ML models significantly improved the accuracy of core inflation forecasts in Japan.
• Predictor Importance: Household inflation expectations, inbound tourism, exchange rates, and the output gap were crucial predictors.
• Model Comparison: LASSO and Elastic Net models outperformed traditional benchmarks.

By leveraging ML techniques, this study provides a robust framework for enhancing inflation forecasting in the face of complex and evolving economic conditions.