Univariate time series forecasting: comparing ARIMA & LSTM neural network to the random walk benchmark for exchange rates

The difficulty of forecasting Exchange Rates has been a longstanding problem for economists and data analysts around the world. Nevertheless, a model that could produce accurate forecasts and outperform the random walk (RW) benchmark would be beneficial to policymakers and investors as it might help mitigate the effects of inflation, thus, having a real impact on the economic perspective. The objective of this paper is to develop and analyze the results of an ARIMA and LSTM models and determine if univariate time series models can show an improved accuracy at a 5-day and 60-day time horizon compared to the driftless random walk (DRW) model, which is the proposed benchmark in this study. In order to perform this analysis, daily exchange rate data for the currency pair USD/EUR was retrieved from the United States Board of Governors of the Federal Reserve System download data program, and later cleaned and manipulated using Python to produce forecasts for each of the models. The predictive accuracy was calculated, and the errors were measured with the MAPE, RMSE, and MSE metrics. Among the three models tested, I concluded that both the LSTM and ARIMA models outperformed the DRW benchmark at the 60-day horizon, which adds evidence of the suitability of autoregressive and machine learning univariate models when forecasting exchange rates, and of their potentially superior performance compared to classical models based on economic fundamentals, which have traditionally failed to outperform the RW benchmark. On the contrary, in the 5-day horizon, the DRW model showed the highest predictive accuracy, thus supporting the existing literature in the difficulty of forecasting exchange rates at short horizons.