Mining profitable high frequency pairs trading forex signal using copula and deep neural network

This work extends the application of recently proposed Mispriced Index (MI), which is based on the conditional probability modeled by copula, to handle high frequency data. It is addressed that the original method (cumulative sum of MI) for handling daily data fails to capture the relationships inherited in the high frequency data. The use of more sophisticated nonlinear models including CART, Logistic regression and Neural network are investigated. The k-fold cross validation strategy is employed to alleviate the problem of data selection bias and provide more reliable evaluations. The empirical results show that MI provides useful information for Neuralnetworkbut not CART and Logistic regression. Significant improvements in profitability, Recall rate, F-measure, ROC index and KS statistics are observed. The modeling performance of Neural networks with 3 to 12 layers are addressed.