Machine learning can produce better forecasts – RBI paper

The use of machine learning (ML) can produce forecasts that are more accurate than standard statistical methods, research published by the Reserve Bank of India finds.

In the working paper, authors Bhanu Pratap and Shovon Sengupta test a broad range of standard statistical forecasting techniques and several different types of ML algorithms, including decision trees and neural networks. They focus on forecasting CPI inflation and its components.

In their test, ML-based models “generally outperformed” standard methods in terms of out-of-sample forecasting performance, they say. However, they add that combining traditional and ML-based forecasting methods can produce even better results.

Such combined forecasts “provided better forecasting accuracy over individual methods in almost all cases”, and worked best when forecasts were combined using a simple average rather than “complex” weightings, they say.

Lastly, the authors compare forecasts based solely on headline inflation, and forecasts made by aggregating the component parts of the headline figure. They find that forecasting inflation directly produces better results.