Football Predictions Based on Time Series with Granular Event Segmentation

Abstract

The method of data quality management in mining fuzzy relations from time series of sports data is proposed. The fuzzy relation-based model of time series with granular event segmentation is developed. Correction of the available expert and experimental information is carried out at the stage of structural tuning of the fuzzy knowledge base. To improve the data set, it is advisable to extract events or specific areas of interest from the time series, rather than analyze the time series as a whole. Experts identify significant events in the form of trends and patterns at the stage of time series segmentation. At the stage of time series granulation, the events in the form of collection of time windows are transformed into fuzzy terms. Expert relationships “fuzzy events – match outcome” are subject to further correction in order to improve the predictive accuracy at the cost of removal of different types of biases in expert judgments. Biases in expert judgements (favorite/outsider bias, home team bias) can be revealed in paired comparative assessments chosen by an expert according to the 9-mark Saaty’s scale. The genetic algorithm is used for off-line finding the area of unbiased assessments with further adaptive correction of paired comparisons matrices embedded into the neural network. As a result, predictive ability of the method outperforms bookmakers’ predictions and, hence, provides the profitability against market odds.