Pattern matching in time series using combination of neural network and rule based approach

Recognizing various meaningful patterns from stock market time series data is getting tremendous attention among researcher during the recent years. Much work has been devoted to pattern discovery from stock market time series data using template based approaches and rule based approaches but not much has attempted to combine the power of any of these approaches with the prediction capability of neural network. We propose here a new novel hybrid pattern-matching algorithm. We combine neural network with rule based approach using variable size sliding window. We focus not only to find regular stock market time series pattern but also for better understanding of the actual stock market, define composite pattern (i.e. composition of approximate simple regular pattern). Specifically, we propose here to model time series data using simple regular pattern and composite pattern simultaneously. Thus, instead of finding isolated simple regular patterns, or predicting the next time series value based on the pattern in the most recent time window, we focus on explaining the relationships between the patterns with the help of composite patterns.