Financial time series forecasting certainly is the case of a predictive modeling process with many challenges, mainly because the temporal structure of the data. Genetic programming, as a particular variation of genetic algorithms, can be used to as a feature engineering, importance and selection process all at once, it can provide highly interpretable symbolic features that have low colinearity among them and yet high correlation with a target variable. We present the use of such method for generating symbolic features from endogenous linear and autoregressive variables, along with a Multi-Layer Perceptron, to construct a binary predictor for the price of Continuous Future Contracts of the Usd/Mxn intra-day exchange rate. The proposition of this work is three fold, first is stated a variation to formulate the classical regression problem of forecasting a continuous value, into a classification problem of forecasting a discrete and binary value, also, in order to address the feature engineering step, the use of Genetic Programming is proposed for producing non linear variables highly correlated with a target and highly uncorrelated with each other, and finally, variations on the performance metrics and Folds of data to perform the training process are implemented. The results are presented for a Logistic regression and a Multi-Layer Perceptron applied to 6 years of historical prices for the UsdMxn Financial Future contract.
