Nowcasting of India's GDP using dynamic factor model: Optimising the results Prakash Anupam1,*, Bhowmick Chaitali2, Thakur Ishu3 1Director, Department of Economic & Policy Research, Reserve Bank of India Mumbai, India 2Manager, Department of Economic & Policy Research, Reserve Bank of India Mumbai, India 3Manager, Department of Economic & Policy Research, Reserve Bank of India Mumbai, India *Corresponding author email id: anupamprakash@rbi.org.in
JEL Classification: C01, C13, C32, C53, C55 Online published on 17 May, 2022. Abstract Monetary policy decisions are based on assessment of current and future economic conditions. Because most macroeconomic data particularly, gross domestic product (GDP) are released with a lag and are subsequently revised, both forecasting and nowcasting are important tasks for central banks. Among the various methods of nowcasting, the dynamic factor models (DFMs) have the advantage that they involve dimension reduction technique to effectively draw real-time information content from large data set with ‘jagged edges’ arising due to varying data release date. Though there are other studies on DFM nowcast for India's GDP, this article uses DFM to forecast real GDP growth published at quarterly frequency based on the estimate of economic cycle represented by the monthly frequency input variables. Within our chosen set of indicators, we entirely cover those used by National Statistical Office (NSO) for their quarterly GDP estimates. Based on the criteria of statistical relationship and timeliness of the indicators, we have specified three alternative model specifications: a 25-indicator model, a 14-indicator model and a 10-indicator model for nowcasting GDP growth. To exploit the new information available with each new release of indicators, three vintages have been prepared to obtain nowcast surrounding a quarter at different points of time. The first vintage (v1) is set at the end of the day of the quarter-ending month (t), second vintage (v2) at a month after (t+30 days) and the third vintage (v3) is set at after 45 days (t+45), when all the input indicators become available. Furthermore, we have applied two alternative estimation techniques: (i) the two-stage estimation method and (ii) the expectation–maximisation (EM) algorithm based on maximumlikelihood method, both of which are repeated across the three vintages. The out of sample root mean square error (RMSE) for the DFM models estimated using EM turned out to be lower than that of SARIMA, while it is higher in case of the DFM estimated using the two-stage method. Also, RMSE improves over successive vintages highlighting the contribution of larger information set. Top Keywords Nowcasting, Short-term prediction, Dynamic factor model, Highfrequency indicator, Monetary policy. Top |