LSTM Variants Comparison for Exchange Rate IDR/USD Forecasting with Rolling Window Cross Validation

The exchange rate is something that is very important in the economy and greatly influences the decisions of global economic actors in making business decisions. This reason, an appropriate data processing model is needed. Many models in the previous research as references. We need model with architectures and suitability for handling sequential data, capturing temporal dependencies, exploring model complexity, and providing a realistic evaluation of predictive performance in time series forecasting scenarios. finally, This research compared of the LSTM prediction models will be carried out, namely vanilla LSTM, stacked LSTM, and bidirectional LSTM by also using a cross-test method called rolling windows cross validation. This method divides the data into contiguous sections, also known as windows, and then cross-validates sequentially on each window. In each window, the model is trained and evaluated in this series of iterations and all evaluation results from these iterations are averaged. The use of this cross-test method to produce a more accurate and reliable evaluation. The results showed that vanilla LSTM produced the best evaluation scores based on existing evaluation metrics, such as RMSE, MSE, MAE and MAPE..