Scholarlink

<p>Accurate market value forecasting is essential for investors, financial analysts, and policymakers due to the inherent volatility and complexity of financial markets. This study employs hierarchical time series (HTS) forecasting to predict the market value of the DJIA index across three levels: overall market value, sector-level market value, and individual stock market value. To capture sequential dependencies in hierarchical financial data, deep learning neural networks—including recurrent neural networks (RNN), long short-term memory (LSTM), and gated recurrent units (GRU)-were utilized for base forecasts. To ensure forecast coherence, various reconciliation methods were applied, including Bottom-Up (BU), Top-Down (Fp), Middle-Out (Fp), MinT (Ols), MinT (Wls-struct), and Optimal Combination (Ols). The effectiveness of these reconciliation approaches was evaluated using performance metrics such as average root mean squared error (AvgRMSE), average mean absolute percentage error (AvgMAPE), average relative RMSE (AvgRelRMSE), and average relative MAPE (AveRelMAPE). The findings indicate that reconciliation methods significantly improve forecast accuracy. Among them, MinT (WLS-struct) consistently achieved the lowest metrics across multiple forecast horizons, including short-term (115 days) and long-term (230 days) predictions for all deep learning 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

Deep Learning Applications in Hierarchical Time Series Forecasting of Market Value

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • John Kamwele Mutinda,
  • Li Yong

摘要

Accurate market value forecasting is essential for investors, financial analysts, and policymakers due to the inherent volatility and complexity of financial markets. This study employs hierarchical time series (HTS) forecasting to predict the market value of the DJIA index across three levels: overall market value, sector-level market value, and individual stock market value. To capture sequential dependencies in hierarchical financial data, deep learning neural networks—including recurrent neural networks (RNN), long short-term memory (LSTM), and gated recurrent units (GRU)-were utilized for base forecasts. To ensure forecast coherence, various reconciliation methods were applied, including Bottom-Up (BU), Top-Down (Fp), Middle-Out (Fp), MinT (Ols), MinT (Wls-struct), and Optimal Combination (Ols). The effectiveness of these reconciliation approaches was evaluated using performance metrics such as average root mean squared error (AvgRMSE), average mean absolute percentage error (AvgMAPE), average relative RMSE (AvgRelRMSE), and average relative MAPE (AveRelMAPE). The findings indicate that reconciliation methods significantly improve forecast accuracy. Among them, MinT (WLS-struct) consistently achieved the lowest metrics across multiple forecast horizons, including short-term (115 days) and long-term (230 days) predictions for all deep learning models.