Decomposition of the training set in the task of forecasting the price of securities

Basic fundamentals of setting the classic problem of forecasting the exchange rate of securities is presented. A training sample was built, consisting of the opening price of the trading session, the maximum and minimum prices and the closing price of the trading session. To build the forecast, the closing prices of the session for 4 stages (trading day) ahead were used. Wolfram Mathematica, a modern machine learning environment for artificial intelligence systems, was chosen as a tool. An optimal structure of a multilayer artificial neural network is built, containing three hidden layers and two data normalization layers. The analytical comparison of the results of predicting the price of securities using only the training set is made, followed by checking on the test set with the version of the complete decomposition of the training set into the training, test and test sets. Evidence has been obtained that the use of a test set to control the accuracy of training at each training step (round) can significantly increase the accuracy of prediction. The training curves for the studied variants of the training set decomposition and the values of the predictive models quality assessment parameters are given.