Temporal pattern matching for the prediction of stock prices

Time series data poses a significant variation to the traditional segmentation techniques of data mining because the observation is derived from multiple instances of the same underlying record. Additionally, the standard segmentation methods employed in traditional clustering require instances to be classified exactly by attaching an event to a specific cluster at the exclusion of other clusters. This paper is an investigation into the predictive power of the clustering technique on stock market data and its ability to provide stock predictions that can be utilised in strategies that outperform the underlying market. This uses a brute force approach to the prediction of stock prices based on the formation of a cluster around the query sequence. The prediction is then applied in a model designed to capitalise on the derived prediction. The predictive accuracy of minimum distance clusters produced promising results with a prediction error incorporated into the forecast strategy.