Decision-making Method of Futures Trading Using Dictionary-based Early Classification of Time Series

The purpose of early classification of time series is to predict the class label of time series in advance when time series has not been collected completely, which is meaningful in financial fields with high timeliness requirements. Current financial analysis techniques, such as methods based on the Support Vector Machine and Naive Bayes, need to analyze complete data to get results, which may delay managers to supervise the market. Therefore, we propose decision-making method of futures trading using dictionary-based early classification of time series. Specifically, we train a group of basic classifiers under different timestamps. The classifier extract subsequences along the sliding window to construct the bag-of-pattern, and then use the logistic regression model for classification. In addition, considering that the main task of early classification of time series is to determine the earliest time of reliable classification. Thus, based on the idea of dynamic decision fusion, we combine the number of classifiers, prediction results of different classifiers, and the conflict function value between earliness and accuracy of results and select the best number of classifiers and the threshold of reliability, which determine the time of reliable output. Consequently, we obtain an algorithm for finding the earliest time of reliable classification. Experimental results on different futures datasets show that, compared with the current popular financial analysis technology, in the aspect of earliness, we use early classification of time series to classify the futures data only by seeing about 60% of length of the complete futures data, which helps the manager of financial regulatory authorities to start making decisions about 40% earlier, leaving more time for judging and guiding decisions. In terms of accuracy, our method has achieved better performance.