Parameter Tuning of Neural Network for Financial Time Series Forecasting

One of the most challengeable problems in pattern recognition domain is financial time series forecasting which aims to exactly estimate the cost value variations of a particular object in future. One of the best well-known financial time series prediction methods is Neural Network (NN) but it suffers from parameter tuning such as number of neuron in hidden layer, learning rate and number of periods that should be forecasted. To solve the problem, this paper proposes a new metaheuristic-based parameter tuning scheme which is based on Harmony Search (HS). To improve the exploration and exploitation rates of HS, the control parameters of HS are adapted during the generations. Evaluation of the proposed method on several financial times series datasets shows the efficiency of the improved HS on parameter setting of NN for time series prediction.