Scholarlink

<p>The ability to accurately predict mutual fund performance is vital for portfolio management and investment decision-making. In this paper, historical fund-level data are combined with macroeconomic indicators into a composite forecasting model that uses three adapted variants of XGBoost and a dynamic combination neural (DCN) meta-model to improve predictive and risk-adjusted performance. The study utilized three distinct XGBoost variants: Yearly Layered XGBoost (YL XGBoost), Multi-Layered Classifier XGBoost (MLC XGBoost), and Optimized Weight XGBoost (OW XGBoost). They were incorporated in the DCN meta model. The inputs were fund-specific attributes and macroeconomic variables, enabling this model to capture both internal fund dynamics and external market patterns. Model performance was measured using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), Sharpe ratio, Treynor ratio, Sortino ratio, maximum drawdown, and coefficient of determination (R²) after preprocessing and feature engineering. Statistical significance was assessed at <i>p</i> = 0.05, and the results were compared with those of previous forecasting models. All three base models demonstrated effective predictive capabilities: OW-XGBoost (RMSE = 0.0055; MAE = 0.0237; Sharpe = 0.0468), YL-XGBoost (RMSE = 0.0048; MAE = 0.0054; Sharpe = 0.0417), and MLC-XGBoost (RMSE = 0.0053; MAE = 0.0065; Sharpe = 0.0472). These base models were stacked into the DCN Meta Model, which achieved significant improvements in RMSE = 0.0983, MAE = 0.2195, and MAPE = 11.75%. It has provided the best Sharpe ratio (1.128), the best Treynor ratio (5.5674), the best Sortino ratio (2.18), and the minimum maximum drawdown (9.87%). Statistical comparisons validated the superiority of the meta model (<i>p</i> &lt; 0.05), and it also outperformed past forecasting techniques and models, such as Bayesian optimization and Random Forests, especially in risk-adjusted models.</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

XGBoost Algorithm-based Model for Predicting Mutual Fund Returns

  • 期刊论文

  • 发表日期：

    2026年1月10日

查看全文
  • Xiaojie Li,
  • Lei Yang,
  • Cheng Zha,
  • Yong Xu

摘要

The ability to accurately predict mutual fund performance is vital for portfolio management and investment decision-making. In this paper, historical fund-level data are combined with macroeconomic indicators into a composite forecasting model that uses three adapted variants of XGBoost and a dynamic combination neural (DCN) meta-model to improve predictive and risk-adjusted performance. The study utilized three distinct XGBoost variants: Yearly Layered XGBoost (YL XGBoost), Multi-Layered Classifier XGBoost (MLC XGBoost), and Optimized Weight XGBoost (OW XGBoost). They were incorporated in the DCN meta model. The inputs were fund-specific attributes and macroeconomic variables, enabling this model to capture both internal fund dynamics and external market patterns. Model performance was measured using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), Sharpe ratio, Treynor ratio, Sortino ratio, maximum drawdown, and coefficient of determination (R²) after preprocessing and feature engineering. Statistical significance was assessed at p = 0.05, and the results were compared with those of previous forecasting models. All three base models demonstrated effective predictive capabilities: OW-XGBoost (RMSE = 0.0055; MAE = 0.0237; Sharpe = 0.0468), YL-XGBoost (RMSE = 0.0048; MAE = 0.0054; Sharpe = 0.0417), and MLC-XGBoost (RMSE = 0.0053; MAE = 0.0065; Sharpe = 0.0472). These base models were stacked into the DCN Meta Model, which achieved significant improvements in RMSE = 0.0983, MAE = 0.2195, and MAPE = 11.75%. It has provided the best Sharpe ratio (1.128), the best Treynor ratio (5.5674), the best Sortino ratio (2.18), and the minimum maximum drawdown (9.87%). Statistical comparisons validated the superiority of the meta model (p < 0.05), and it also outperformed past forecasting techniques and models, such as Bayesian optimization and Random Forests, especially in risk-adjusted models.