Scholarlink

<p>Stock return predictability has been one of the most studied fields in finance. This research examines SPDR S&amp;P500 ETF Trust—SPY’s—daily data from December 3, 2007, to May 20, 2025, employing machine learning (including Logistic Regression, Ridge Classifier, CatBoost Classifier, Decision Tree Classifier, Naïve Bayes, and AdaBoost Classifier) and deep learning techniques (Dense, LSTM, GRU, CNN, and Hybrid-CNN models) under varying risk aversion thresholds. The SPY is a fund aimed at mirroring the price and yield performance of the S&amp;P500 Index, before fees, to provide investors access to the US equity market via one diversified investment. The analysis is extended to a weekly and monthly basis for further comprehension. We introduce a dynamic, rolling-window-based thresholding method for defining the binary target variable, as well as a quintile-optimized threshold for calibrating classification outputs. The analysis is extended beyond daily data to weekly and monthly frequencies to evaluate robustness across investment horizons. Results consistently highlight the superior generalization ability of GRU and CNN–GRU models, especially in out-of-sample testing, where they achieve high F1-Scores and balanced precision–recall tradeoffs. While ensemble tree-based methods such as CatBoost and AdaBoost achieve high accuracy in training, they show signs of overfitting in test sets, particularly at lower data frequencies. Overall, the findings support the robustness of recurrent and hybrid deep learning architectures in adapting to financial return dynamics across time scales, while reaffirming the utility of classical machine learning models under proper calibration.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Stock Return Predictability on S&P 500 ETF: A Deep Learning Approach

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Héctor J. Hortúa,
  • Andrés Mora-Valencia,
  • Esteban Vanegas

摘要

Stock return predictability has been one of the most studied fields in finance. This research examines SPDR S&P500 ETF Trust—SPY’s—daily data from December 3, 2007, to May 20, 2025, employing machine learning (including Logistic Regression, Ridge Classifier, CatBoost Classifier, Decision Tree Classifier, Naïve Bayes, and AdaBoost Classifier) and deep learning techniques (Dense, LSTM, GRU, CNN, and Hybrid-CNN models) under varying risk aversion thresholds. The SPY is a fund aimed at mirroring the price and yield performance of the S&P500 Index, before fees, to provide investors access to the US equity market via one diversified investment. The analysis is extended to a weekly and monthly basis for further comprehension. We introduce a dynamic, rolling-window-based thresholding method for defining the binary target variable, as well as a quintile-optimized threshold for calibrating classification outputs. The analysis is extended beyond daily data to weekly and monthly frequencies to evaluate robustness across investment horizons. Results consistently highlight the superior generalization ability of GRU and CNN–GRU models, especially in out-of-sample testing, where they achieve high F1-Scores and balanced precision–recall tradeoffs. While ensemble tree-based methods such as CatBoost and AdaBoost achieve high accuracy in training, they show signs of overfitting in test sets, particularly at lower data frequencies. Overall, the findings support the robustness of recurrent and hybrid deep learning architectures in adapting to financial return dynamics across time scales, while reaffirming the utility of classical machine learning models under proper calibration.