As a major symbolic representation method that has been widely used in time series data mining, Symbolic Aggregate approXimation (SAX) uses the mean value of a segment as the symbol. However, the SAX representation ignores the trend of the value change in the segment, which may cause incorrect classification in some cases, because it cannot distinguish different time series with different trends but the same average value symbol. In this paper, we propose an improved symbolic representation by integrating SAX with the least squares method to describe the time series' mean value and trend information. By comparing the classifiers using the original SAX, two improved SAX representations and another two classifiers that are highly representative and competitive for short time series classification, the results show that the error rate of the classifier that uses our representation is lower than that of those five classifiers on their own in most datasets.