Expanding RBI's toolkit

Machine learning can improve inflation forecasting

The importance of accurate inflation forecasting for monetary policy cannot be overstated and central banks across the world have made big errors in recent years. Besides central banks, inflation forecasts also hold importance for households and firms, which negotiate wage-price contracts. In this context, economists at the Reserve Bank of India (RBI) recently came up with a research paper, comparing the accuracy of machine learning (ML) vis-à-vis traditional econometric models in forecasting inflation in the country. Using quarterly data from the second quarter of 1996 to the first quarter of 2022, the study notes that ML techniques outperform econometric models for both pre-pandemic and post-pandemic periods. ML techniques are superior to non-ML alternatives for longer forecast horizons and in the presence of structural breaks. The forecasts were then compared with actual year-on-year headline inflation and median inflation forecasts of professional forecasters — which are representative of industry inflation expectations.

The relationship between inflation and its determinants has become increasingly non-linear and complex over time, making forecasting more challenging. Unanticipated macroeconomic shocks such as the pandemic further distorted the predictive power of traditional forecasting models, which rely on historical trends and patterns. ML techniques, on the other hand, generate lower forecast errors and are associated with increased efficiency. The headline inflation rate in India remained above the 4 per cent target for 49 of the 60 months between November 2018 and October 2023. It exceeded the upper tolerance band of 6 per cent in 27 months. Such a large variation has led to episodes of persistent and high forecast errors. In 2019, RBI economists attributed this to unanticipated shocks emanating from prices of food items. Cross-country evidence is suggestive of a positive correlation among forecast errors with the share of food items in the consumer price index (CPI) basket. This fact is consistent with India’s experience where the share of food items in the CPI basket is 46 per cent.

The central bank has been making sustained efforts to improve its inflation-forecasting performance. Back in 2021, the RBI revised its inflation-forecasting model to capture how fiscal and monetary policies interact with real-economy elements. Dubbed “Quarterly Projection Model 2.0”, it was an improvement over the previous version, which often overestimated upside risks to inflation. In the past decade or so, the proliferation of the internet has increased central banks’ exposure to data, based on micro-transactions among agents (e-commerce and credit-card transactions), public statistics, and financial-market data. Macroeconomic and financial market big data, however, exhibits noise, nonlinear patterns, or temporal dependencies, rendering traditional econometric methods insufficient to analyse them. The use of artificial intelligence (AI) and ML has gained importance in this process.

Central banks all over the world are now deploying ML-based techniques owing to their ability to deal with larger and new sources of information in a more automated way. A few months back, for instance, the RBI selected two global consultancy firms to upscale its advanced analytics and develop AI-ML systems for its supervisory functions. There is a case for expanding the use of AI-ML to areas like systemic risk detection, financial crime management, creditworthiness scoring, and forecasting business cycles. Judicious and prudent use of new tools to improve inflation-forecasting outcomes should help in conducting monetary policy and anchoring inflation expectations.