Machine learning algorithms for quarterly GDP forecasting: A performance evaluation

Machine learning (ML) algorithms have emerged as potential competitors to the conventional econometric models in the empirical forecasting domain. Using the quarterly real GDP series of Indian economy covering the period 1996Q1– 2020Q4, we investigated in this study whether ML models prove themselves superior over the familiar econometric models. Among the competing ML models, we applied (a) the Facebook's prophet algorithm and (b) the Pattern Sequencebased Forecasting (PSF) algorithm. Among the conventional econometric models, we applied Auto-Regressive Integrated Moving Average (ARIMA), while also taking into account the seasonality in the present series, that is, Seasonal Auto-Regressive Integrated Moving Average (SARIMA). The Augmented Dickey–Fuller (ADF), the Phillips–Perron (PP), the Dickey–Fuller Generalized Least Squares (DF-GLS) and Kwiatkowski–Phillips–Schmidt–Shin (KPSS) unit root tests were employed to investigate stationarity in the series. Forecasts were made for the period 2021Q1–25Q4, where the year 2025 is targeted to reach a US$ 5 trillion economy.

Results showed that the SARIMA model generated estimates closer to the actual GDP over the ML model prophet in the trained series. This was followed by the prophet and ARIMA, while the other ML model PSF lagged behind. Hence, the results in this study add to the literature claiming ‘mixed evidences’ in evaluating the performance of ML-based algorithms in time series forecasting. Adding a number of other econometric and ML-based models is suggested while probing this evaluation further.