Abstract
Predicting the stock market has become an increasingly interesting research area for both researchers and investors, and many prediction models have been proposed. In these models, feature selection techniques are used to pre-process the raw data and remove noise. In this paper, a prediction model is constructed to forecast stock market behavior with the aid of independent component analysis, canonical correlation analysis, and a support vector machine. First, two types of features are extracted from the historical closing prices and 39 technical variables obtained by independent component analysis. Second, a canonical correlation analysis method is utilized to combine the two types of features and extract intrinsic features to improve the performance of the prediction model. Finally, a support vector machine is applied to forecast the next day's closing price. The proposed model is applied to the Shanghai stock market index and the Dow Jones index, and experimental results show that the proposed model performs better in the area of prediction than other two similar models.
Introduction
Predicting the stock market is notoriously difficult due to its non-linear, noisy, and chaotic nature. Most existing models fall into one of two camps: auto-regression models, which look only at historical price data, and multi-variable models, which utilize technical indicators like moving averages or the Relative Strength Index (RSI).
The authors argue that both approaches are incomplete. Historical prices provide market "momentum," while technical indicators provide "context." However, simply combining both types of data into a machine learning model creates a high-dimensional mess of redundant information and noise. The challenge lies in effectively "fusing" these two different types of information into a single, clean input for a predictor.
Methodology
The researchers developed a three-stage pipeline called ICA-CCA-SVR, designed to reduce noise and extract the most predictive features from complex financial data.
- Feature Extraction (ICA): Independent Component Analysis (ICA) is used to process both raw historical prices and 39 technical indicators. Unlike Principal Component Analysis (PCA), ICA identifies statistically independent components, making it better suited for non-Gaussian financial data.
- Feature Fusion (CCA): This core innovation applies Canonical Correlation Analysis (CCA) to combine price-based and indicator-based features. CCA identifies the "intrinsic" relationship between these sets, removing redundant noise and retaining the most predictive "union features."
- Prediction (SVR): These fused features are fed into a Support Vector Regression (SVR) model. SVR is chosen for its ability to handle non-linear data and its use of structural risk minimization to avoid overfitting.
Data Preparation and Experimental Setup
The model was evaluated using two major real-world datasets from January 2003 to December 2005: the Shanghai Stock Market Index (representing emerging markets) and the Dow Jones Index (representing mature markets).
To prepare the data for the SVR, a sliding window approach was used to generate training and testing samples from the historical trade data.
Technical indicators included 39 variables such as Moving Averages (MA6, MA12), BIAS, MACD, Stochastic %K/%D, and the Psychological Line (PSY).
Key Results and Insights
The proposed ICA-CCA-SVR model was compared against two benchmarks: AICA-SVR (price only) and MICA-SVR (indicators only).
- Higher Accuracy: On the Shanghai Index, the proposed model achieved a correlation coefficient ($r$) of 0.9517, significantly higher than the price-only (0.8174) and indicator-only (0.9168) models.
- Lower Error: The Mean Absolute Percentage Error (MAPE) for the proposed model was just 0.011 for the Shanghai Index and 0.0056 for the Dow Jones, consistently lower than the benchmarks.
- Stability: The model showed much more stable predictive performance as feature dimensionality increased, whereas benchmark models' performance fluctuated.
The study highlights that historical prices and technical indicators are "correlating and complementary." While one indicates past events, the other suggests underlying strength. Effective fusion via CCA prevents the "curse of dimensionality" that often occurs when simply concatenating feature vectors.
Conclusion
While the ICA-CCA-SVR model demonstrates superior performance, it does have limitations. It requires a certain level of data complexity (dimensionality > 3-6) to realize the benefits of fusion. Furthermore, the slightly lower accuracy on the Dow Jones Index suggests that mature markets may contain more complex noise requiring further refinement. Future work may explore "parallel" feature fusion methods to further enhance predictive accuracy.
Summary updated