Scalable Financial Time Series Pattern Recognition with Ratio Based Data Representation

Financial time series data are highly dimensional by nature, thus price pattern recognition is fundamentally challenging as similar price patterns might transpire in multiple timeframes at a very different scale. In the mining of hidden price patterns, the complication exacerbates as the extraction of price pattern samples requires a search on all the possible lengths since the samples of the price patterns are unknown in advance. As such, this study proposed a ratio-based data representation by applying a dynamic ratio that corresponds to the fluctuation in the financial time series price patterns. This optimizes the dimension reduction of the original time series data without missing the representation of the financial time series price pattern, at the same time significantly reduces the search space of the sliding window search for hidden financial time series price patterns with varying lengths.