This work proposes a hybrid LSTM-GARCH model to forecast the volatility of the USD-COP exchange rate, known as tasa representativa del mercado (TRM). This model is a LSTM recurrent neural network which includes coefficients of time series models GARCH, EGARCH and EWMA for the TRM as explanatory variables, according to the framework from Kim et al. 2018. Both GARCH and LSTM models are estimated with historical data from 2008 to June 2018, and the estimated forecasts are contrasted with data from July 2018 to July 2019. The acquired forecasts are compared using various linear and nonlinear error measures. The inclusion of coefficients from GARCH and EGARCH models is shown to improve the accuracy of hybrid LSTM-GARCH model predictions compared to a standard LSTM model.
