It’s Not Just You, Artificial Intelligence May Even Help India’s Central Bank Become Smarter
The Reserve Bank of India’s inability to forecast a sharp drop in inflation levels over the last two years has drawn much criticism. The miss has prompted much debate and even ‘introspection’ within the central bank.
As it turns out, the central bank may have been able to avoid some of this embarrassment had it used machine learning algorithms to predict headline inflation rather than rely on its traditional statistical models.
According to a working paper released by the central bank on Tuesday, RBI researchers found that machine learning techniques proved to be more accurate in forecasting retail inflation in India.
Machine Learning vs Statistical Models
The research focused on CPI inflation and three of its components as part of the study. The three components apart from headline inflation were food and beverages inflation, fuel and light inflation and core inflation.
In judging accuracy, the research used seven machine learning models and four statistical models.
The result? All seven machine learning models outperformed three of the four statistical models including the commonly-used ‘Random Walk’ model.
One statistical model, however, managed to outperform the machine learning techniques. The statistical model that outperformed is one known as ‘’SARIMA’ or the ‘Seasonal Autoregressive Integrated Moving Average Model’.
The difference in projections was stark.
Each of the seven machine learning techniques predicted that inflation would fall from 4-5 percent in July 2018 to a range of 2-3 percent by December 2018. Three of the four statistical models predicted that inflation may fall to between 2-3 percent.
The inflation predicted by machine learning models was closer to actual inflation levels, which fell from 4.17 percent in July 2018 to 2.19 percent in December 2018.
Headline Inflation vs Inflation Components
The study also found that directly forecasting headline inflation proved to be more accurate than forecasting sub-sets of inflation, such as food or fuel inflation, and then deriving headline inflation.
Traditional econometric models focus on producing good coefficient estimates to explain the underlying relationship between two or more variables, explained the paper. Instead, machine learning directly produces predictions of one variable from another by fitting complex and flexible functional forms on data.
The study also found that simple average-based forecast combinations outperformed complex weighted average combination of forecasts.
Are Central Banks Using Artificial Intelligence?
Accuracy in forecasting inflation is critical for central bankers, who use price movements as a key input into interest rate decisions. This is particularly true for inflation-targeting central banks, including the Reserve Bank of India, which follows a flexible inflation targeting framework.
So are central banks using artificial intelligence to improve their forecasts and decision making?
According to the study, applications of big data and machine learning-based analytics are increasingly being used by central banks. Large central banks have established their own data analytics teams with the aim to tackle key issues related to market regulation, supervision, surveillance, risk management and monetary policy.
“For central banks, such advancements call for addition of newer tools to their analytical toolkit – ones which are adept at handling large volume, high-frequency data,” the study said.