Modeling and Prediction of Time-Series-A Case Study with Forex Data

Time series data reveals dynamic behavior of systems. A few real life examples are traffic flow, amount of rainfall, usage of electricity, share values, Forex rate etc.. Depending on the complexity of the system dynamics, algorithms differ to model the time series data accurately, so that the created model can be used for interpolation and more commonly extrapolation or prediction. For example, AR model performs well in stationary time series, but for non-stationary, it cannot capture the non-linear dynamics. In this research, we use Forex rate data, and experimented with various algorithms to capture the dynamics of the data. The success of the model is evaluated by accuracy in prediction.In our experiments, we applied two state-of-the-art models -Support Vector Regression (SVR) and Recurrent Neural Network (RNN). The target of the experiment is the prediction of longer future by recursion (feeding back predicted value to input for the next step prediction). The result shows that RNN with proper Long Short-Term Memory (LSTM) has better performance in predicting longer future.