Abstract
Stock index tracking is a critical task in passive investment, yet it is complicated by the high-dimensional, non-linear, and noisy nature of financial data. This paper proposes a hybrid framework, LSTM-DGMDH, to address these challenges. The methodology integrates an attention-based Long Short-Term Memory (LSTM) network for data preprocessing with a Deep Evolutionary Group Method of Data Handling (DGMDH) neural network for predictive modeling. By introducing a novel Tracking Error External Criterion (TEEC), the model optimizes portfolio allocation based on yield differences rather than simple error metrics. Experimental results on the CSI 300 index demonstrate that the proposed approach significantly enhances tracking accuracy and reliability compared to traditional regression models.
Introduction
Index tracking aims to construct a portfolio of stocks that replicates the performance of a target market index. In modern finance, this is a cornerstone of passive management. However, stock indices like the CSI 300 are characterized by extreme complexity, including time delays, uncertainty, and high dimensionality. Furthermore, financial datasets often contain missing or abnormal values that can destabilize traditional tracking models.
To overcome these hurdles, researchers have turned to deep learning. However, standard models often struggle with the "black box" nature of deep networks or fail to handle the specific noise inherent in financial time series. The LSTM-DGMDH framework presented here addresses these issues through a two-stage process: sophisticated data imputation and an evolutionary, self-organizing modeling approach.
Methodology
Attention-LSTM Preprocessing
The first stage of the pipeline focuses on cleaning and reconstructing the input data. While standard LSTMs are effective at capturing temporal dependencies, they treat all historical time steps with equal importance in their default state.
To improve this, the authors introduce an attention mechanism. This mechanism assigns dynamic weights to the hidden states of the LSTM. By learning which past observations are most relevant to the current state, the model can more accurately fill in missing values and smooth out noise, creating a robust dataset for the tracking phase.
Deep Evolutionary GMDH (DGMDH)
The core tracking model utilizes the Group Method of Data Handling (GMDH), a self-organizing neural network architecture. Unlike traditional deep networks that require manual tuning of layers and neurons, GMDH "evolves" its structure.
The authors employ a discrete form of the first-order Kolmogorov-Gabor (K-G) polynomial as the reference function. The network starts with simple combinations of input variables and iteratively builds more complex layers. This evolutionary process continues until the model reaches "optimal complexity"—the point where further growth would lead to overfitting.
Tracking Error External Criterion (TEEC)
A key innovation of this work is the Tracking Error External Criterion (TEEC). Traditional external criteria used in GMDH often rely on generic error metrics like Mean Squared Error. However, in index tracking, the primary goal is to match the target yield. TEEC specifically measures the difference between the allocation yield of the constructed portfolio and the target index yield. This domain-specific metric provides better information for selecting the optimal neurons during the DGMDH evolutionary process.
Experimental Evaluation
The model was rigorously tested using data from the CSI 300 index, covering the period from July to November 2021. The researchers developed four distinct datasets to test the model's versatility:
- Clopri102 & Clopri51: Focused on closing prices with varying observation frequencies.
- Yierat102 & Yierat51: Focused on yield rates.
These datasets present a significant challenge: they involve 300 variables (individual stocks) but a relatively small number of temporal observations, making them highly susceptible to the "curse of dimensionality."
Results and Discussion
The LSTM-DGMDH model was compared against five established benchmarks, including Gaussian Process Regression (GPR) and Elastic-Net (NAEN).
- Superior Accuracy: The DGMDH model consistently achieved the lowest Root Mean Square Error (RMSE) and Tracking Error (TE) across all test sets.
- Statistical Dominance: Using Copeland Scoring to aggregate performance across multiple metrics, the proposed methodology emerged as the clear leader, significantly outperforming traditional statistical and machine learning approaches.
- Interpretability: Because the DGMDH model is built upon polynomial functions, it offers a higher degree of mathematical interpretability compared to standard "black-box" deep learning models.
Conclusion
The LSTM-DGMDH framework represents a significant advancement in high-dimensional index tracking. By combining the temporal processing power of Attention-LSTMs with the self-organizing capabilities of evolutionary GMDH networks, the model effectively handles the noise and complexity of financial markets. Future work will focus on extending this methodology to longer time horizons, accounting for transaction costs, and validating the model on global indices such as the S&P 500.
Summary updated