Abstract
The stock market, a cornerstone of the securities industry, presents significant challenges for investors due to its inherent complexity and the multitude of influencing factors. This paper proposes a hybrid prediction model, LASSO-BPNN, designed to forecast stock fluctuation trends. The methodology involves a two-stage process: first, the Least Absolute Shrinkage and Selection Operator (LASSO) method identifies variables with significant impact on stock prices; second, a Back Propagation Neural Network (BPNN) is constructed using these selected variables to predict trends. Empirical testing on a Chinese new energy stock (A-share market) demonstrates that the LASSO-BPNN model significantly improves prediction accuracy compared to standalone neural networks, offering a robust tool for market analysis and investment decision-making.
Introduction
Predicting stock market movements is a notoriously difficult task because the market functions as a complex, non-linear system. Traditional statistical models, such as ARIMA, often fail to capture the intricate relationships between macroeconomic indicators and individual stock performance. While modern Artificial Intelligence (AI) and neural networks offer powerful modeling capabilities, they are susceptible to "overfitting" when provided with excessive or irrelevant input data—a phenomenon often described as "garbage in, garbage out."
The primary challenge addressed in this research is effective variable selection. By determining which economic indicators truly drive a specific stock's price before training a predictive model, we can enhance both the efficiency and the accuracy of the forecast.
Methodology: The LASSO-BPNN Pipeline
To address the limitations of high-dimensional data in neural networks, this paper introduces an integrated pipeline combining statistical regularization with deep learning.
1. Feature Selection via LASSO
The Least Absolute Shrinkage and Selection Operator (LASSO) is employed to refine the set of potential predictors. LASSO performs both variable selection and regularization by applying a penalty to the coefficients of the variables. In this process, the coefficients of less significant variables are shrunk to zero, effectively removing "noise" from the dataset. This ensures that only the most impactful macroeconomic and industry-specific factors are passed to the prediction stage.
2. Trend Prediction via BPNN
The variables prioritized by the LASSO filter serve as inputs for a Back Propagation Neural Network (BPNN). The BPNN is a multi-layer feedforward network that utilizes an error back-propagation algorithm to learn the non-linear mappings between the selected factors and the stock's closing price.
Empirical Analysis
The study focused on a specific new energy stock in the China A-share market (code 601012) using 103 months of data from May 2012 to November 2020.
Variable Selection Results
Initially, nine potential factors were considered: per capita GDP (F1), oil price (F2), market environment (F3), consumer confidence index (F4), consumer price index (F5), sales price (F6), coal production (F7), consumer attitude (F8), and after-sales service (F9).
The LASSO method revealed that several commonly cited indicators, such as GDP and oil prices, were not significant drivers for this specific stock. The final model utilized five key variables: market environment (F3), consumer confidence (F4), sales price (F6), coal production (F7), and consumer attitude (F8). The analysis established a hierarchy of influence: Coal production > Sales price > Market environment > Consumer confidence > Consumer attitude.
Results and Discussion
The performance of the LASSO-BPNN model was compared against a standard BPNN model using data from June 2020 to November 2020 as a test set.
- Accuracy: The LASSO-BPNN model achieved an average relative error of 6.52%, a substantial improvement over the 11.15% error rate of the standalone BPNN.
- Stability: The hybrid model demonstrated greater consistency, with a maximum error of 8.70%, compared to the 15.27% peak error observed in the standard model.
- Directional Forecasting: A critical advantage of the LASSO-BPNN approach was its ability to correctly predict the direction of fluctuations. While the standard BPNN failed to identify a trend reversal between July and August 2020, the proposed model accurately tracked the rising and falling trends throughout the test period.
Conclusion
The LASSO-BPNN model provides a more accurate and stable framework for predicting stock fluctuation trends by effectively filtering irrelevant data before the learning phase. While the results are promising for the new energy sector, future research should expand the dataset to higher frequencies (daily or weekly) and test the model across different industrial sectors to validate its generalizability.
Summary updated