Forecasting Short-Term Volatility of the Indian Rupee to United States Dollar Exchange Rate: A Comparison of a HAR-RV Linear Model and a Random Forest

The value of the Indian rupee against the United States dollar changes every trading day, and how much it moves, its volatility, matters for importers, exporters, investors, and policymakers. This study asks a simple question: when forecasting short-term volatility in the rupee to dollar (INR/USD) exchange rate, does a more complex machine learning model do better than a simple statistical one? Daily spot exchange rates published by the United States Federal Reserve were used to build a dataset covering 2010 to 2016, giving 1,727 trading days after cleaning. From the rates, daily, weekly, and monthly measures of realized volatility were calculated and used to predict volatility over the following five days. Two models were compared: a linear regression in the style of the Heterogeneous Autoregressive Realized Volatility (HAR-RV) model, and a random forest. Both were trained on 2010 to 2015 and tested on 2016, which they never saw during training. Forecasts were scored with four error measures: MAE, MSE, RMSE, and MAPE. The linear model was more accurate than the random forest on every measure, with a lower average error (MAE of 0.00282 against 0.00309) and a lower percentage error (about 66 percent against 76 percent). Both models tracked the average level of volatility but missed sudden spikes. The results suggest that for this task a simple and interpretable model can outperform a more flexible one.