Pattern-Based Feature Extraction for Improved Deep Learning in Financial Time Series Classification

In this paper, the authors introduce a novel feature extraction method based on pattern detection in financial data to enhance the performance of deep learning models for financial time series classification. Existing financial forecasting models often struggle with the inherent volatility and complexity of financial markets, particularly due to their reliance on traditional financial data features which fail to capture intricate price patterns. This research addresses the critical gap in effectively identifying and leveraging these patterns to improve predictive accuracy. By collecting tick-by-tick data from the European Carbon Emission Allowance market and resampling it to a 15-minute timeframe, we developed a method to detect pivotal price patterns and extract relevant features. Integrating these features with traditional open, high, low, close, and volume (OHLCV) data in long-short-term memory (LSTM) and gated recurrent unit (GRU) combined with dense neural networks, our empirical results demonstrate significant improvements in model performance, showcasing enhanced accuracy compared to models using only traditional data features.