The main aim of this study is to compare the accuracy of econometrics and deep learning models in forecasting the exchange rate volatility of the US dollar and Sri Lankan rupee in a multivariate framework. Foreign currency movements in a country mainly arise from export and import transactions recorded in the current account, as well as from Foreign Direct Investments, bank borrowings, and other capital inflows recorded in the capital account. In addition, foreign exchange flows are influenced by economic, political, and financial uncertainties within the country. Accordingly, as crude oil price is the largest import item representing the country’s current account, the 3-Month Dollar Sri Lankan Rupee Forward Rate, All-Share Price Index Returns, Deutscher Aktien Index Returns, Dow Jones Industrial Average Returns and Hang Seng Returns are selected as representations of the capital account via covered interest arbitrage conditions, the S&P 20 Sri Lanka Index and MSI Sri Lanka Index. These represent the relative uncertainty in the Sri Lankan market, where foreign markets use a set of independent variables with the USD/LKR exchange rate as the dependent variable. Daily data was collected from December 2011 to April 2023, and DCC-GARCH, LSTM and DCC-GARCH-LSTM hybrid models were employed for the comparison. According to the model accuracy results measured by MSE, RMSE and MAE, the USD/LKR exchange rate is best predicted using LSTM, followed by DCC-GARCH and the hybrid model. Sri Lankan exchange rates operate under managed conditions and are subject to policy changes, causing non-linear patterns that can be more accurately captured by LSTM models than by GARCH models, as the structure is primarily volatility-driven in such models. Furthermore, DCC-GARCH may not react instantly to sudden changes, but it effectively captures a change that has occurred. Hybrid models combine this slow adjustment with the faster reactions of an LSTM, which can create conflicting signals, making prediction difficult.