Predicting market inflation expectations with news topics and sentiment

TL;DR

This paper introduces a machine learning approach that incorporates news topics and sentiment to improve inflation expectation predictions across multiple countries, highlighting the importance of economic, health, and government topics.

Contribution

It develops a novel predictive framework using narrative-based features and demonstrates their effectiveness over traditional models in forecasting inflation expectations.

Findings

Machine learning models with news features outperform benchmarks.

Logistic Regression and XGBoost achieve top performance.

Economic, health, and government topics are key predictors.

Abstract

This study presents a novel approach to incorporating news topics and their associated sentiment into predictions of breakeven inflation rate (BEIR) movements for eight countries with mature bond markets. We calibrate five classes of machine learning models including narrative-based features for each country, and find that they generally outperform corresponding benchmarks that do not include such features. We find Logistic Regression and XGBoost classifiers to deliver the best performance across countries. We complement these results with a feature importance analysis, showing that economic and financial topics are the key performance drivers in our predictions, with additional contributions from topics related to health and government. We examine cross-country spillover effects of news narrative on BEIR via Graphical Granger Causality and confirm their existence for the US and Germany, while five other countries considered in our study are only influenced by local narrative.