Comparative Analysis of Machine Learning Models for Money Demand Forecasting in the Indian Economy

HSE Economic Journal. 2024. Vol. 28. No. 1. P. 133–158.

Sikhwal S., Sen S.

The study investigates the predictive efficacy of various machine learning methodologies, encompassing Random Forest (RF) regression, Gradient Boosting (GB), Xtreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), Least Absolute Shrinkage and Selection Operator (LASSO) regression, and a deep learning technique, specifically Long Short-Term Memory (LSTM). The benchmark method employed is the autoregressive (AR) model of order 1. With a focus on forecasting money demand for the Indian economy, a crucial component for achieving the Central Bank of India's inflation targeting objective, a comprehensive monthly dataset from 1997 to 2021 is utilized. The obtained results underline the robust predictive capabilities of the employed models concerning both narrow and broad money demand forecasts. By employing a range of evaluation metrics, the study rigorously compares the predictive performance of these models. Using the expanding window cross validation with time series split, the models are cross-validated to ensure accurate forecasts of monetary aggregates. Moreover, the Diebold – Mariano test is utilized to evaluate and compare the quality of forecasts. In particular, the research finds the superiority of LSTM and LASSO in predictive capabilities for narrow and broad money demand, respectively. These findings collectively contribute to enhancing the understanding of money demand prediction, thus facilitating informed decision-making within the realm of monetary policy.

Keywords: Extreme gradient boosting, machine learning models, broad money demand, narrow money demand, Random Forest regression

Publication based on the results of: Institutions and human capital as factors in crisis adaptation (2024)