Currency Exchange Rate prediction using LSTM and Bi-LSTM Model

Abstract

Predicting financial data, whether stock market rates or foreign exchange rate prices, has always been tricky. Financiers all throughout the world are interested in accurate data prediction because of the possible financial rewards. According to Efficient Market theory financial market is random process and impossible to predict accurately. However, researchers are trying different methods to predict market values. Different statistical approaches like Moving Average, Smoothing Average, ARIMA and other models have shown quite better results for prediction. Now with advent of processing power machine learning algorithms have also shown great efficiency. Neural networks are quite popular due to its highest accuracies. In this work we have worked on prediction of currency exchange rates through 2 different by using a modified version of Convolution Neural Networks called Long Short Term memory (LSTM). We generated two models in this experiment, one using LSTM and the other using Bi LSTM. NASQAD Composite (IXIC) is an American Stock index, that we have tried to model and predict using both the models. For modelling we have used historical price closing data. The stock's information was acquired from Yahoo Finance. As an accuracy statistic, we applied MSE. We compared its accuracy to previously published results for the same dataset. We have noticed that models have significantly improved the accuracy. We have also applied these models to Forex data for seven different currency pairs. Daily closing exchange rates were available in the data for AUDUSD, EURUSD, GBPUSD, NZDUSD, CADUSD CHFUSD and JPYUSD. The dataset recorded values of these currencies for each 5 minutes from Oct 2020 to March 2021. We selected one instance of values per day instead of each five-minute value. The dataset for these values were obtained from State Bank of Pakistan for research. Results have shown that our Bi-LSTM Model has far better performance than LSTM model. NASQAD IXIC stock market MSE reported by researches using LSTM is 0.0022 while as proposed LSTM has achieved MSE of 0.0001 which is quite less. Forex Data set has an MSE of 0.00014 for LSTM and 3.25 x 10 ^ -5 for Bi-LSTM model. From our research, we can state that these models are crucial for correct forecasts of time series data