Machine learning can help forecast inflation yields better: RBI Report

Empirical estimates indicate that the utilization of Machine Learning (ML)-based techniques for forecasting inflation yields superior performance gains compared to traditional methods, according to a report by the Reserve Bank of India. Notably, these gains are found to be significantly higher during the post-pandemic period, the report said.

“The empirical results suggest performance gains in using ML-based techniques over traditional ones in forecasting inflation in India over different forecast horizons,” the report said.

The report said that in the post-COVID period, unforeseen fluctuations in both supply and demand led to a substantial increase in mean inflation, accompanied by a reduction in inflation volatility. Consequently, trend inflation, which had consistently declined since 2011, experienced an upward shift. Similar to previous economic crises, the pandemic appears to have potentially introduced structural changes to the inflation process. Notably, a formal structural break test conducted on headline inflation indicates a break coinciding with the onset of the pandemic.

The report said that on average, ML techniques demonstrated superior performance compared to traditional linear models in both pre-pandemic and post-pandemic periods. The heightened performance gains observed with ML techniques for forecasting one-quarter and four-quarters ahead during the post-pandemic period suggest their effectiveness in capturing the increased volatility associated with the pandemic's impact on inflation.

“On average, ML techniques outperformed the traditional linear models for both the pre-pandemic and post-pandemic periods. The performance gains achieved using the ML techniques over the one-quarter ahead and four-quarters ahead forecast horizons were significantly higher in the post-pandemic period, implying that ML techniques may be better at capturing the pandemic time volatility in inflation,” the report said.

When using the Survey of Professional Forecasters (SPF) as the target, a similar trend emerges—forecasts from Deep Learning (DL) models align more closely with SPF median forecasts than those from traditional models, the report said. However, the report also said that considering the inherent challenges of ML models, such as complex structures, over-parameterization, and limited interpretability, it remains essential to regularly assess and compare the forecasting capabilities of ML techniques against traditional models across various sample periods.