Stock trend analysis based on feature rank by partial least squares

Time series analysis, an important domain of machine learning and data mining, has attracted much research interest in the previous decade. Part of the research efforts have focused on analyzing stock data, which is a major application of time series analysis. However stock data is always considered as a unidimensional sequence in recent research lacking of the whole market perspective. This paper regards the stock data as a multi-dimensional time series and proposes a novel method of stock trend analysis based on feature reduction method. By taking advantage of the partial least squares, we rank the dimensions of stock data firstly. Then a feature selection algorithm based on the dimension rank is developed referring to the wrapper algorithm. Finally the algorithm is utilized to analyze the stock trend in both single stock and the whole market perspectives on the real stock data. Experimental results show the potential of the feature reduction based trend analysis method of multi-dimensional series, which performs better to the unidimensional-based methods and obtains the global information from the whole market view.