Data Extension for Stock Market Analysis

Time series of stock prices were subjects of multivarietal statistical analysis. After basic data description obtained by logarithmic differences, several methods of data augmentations were implemented. Several dozen of new variables were computed from sliding window of the differentiated time series using simple statistic tools like first and second moments as well as more complicated statistic, like auto-regression coefficients and residual analysis, followed by an optional quadratic transformation that was further used for data extension. Regularized logistic regression helped in prediction of Buy-Sell Index (BSI) from real stock market data. Various measures of prediction quality were discussed. The regularization gain affected both the number of descriptors and prediction accuracy. The prediction system was optimized on a group of stock series and then cross-validated on another group of stocks.