Scholarlink

<p>This paper evaluates whether <i>forecast ensembles</i> improve out-of-sample predictions of the monthly U.S. equity risk premium, relative to standard single-predictor ARDL regressions. Using thirty macro-financial and technical indicators, we assess the performance of boot-strap bagging (BA), five linear combination schemes, and principal component (PC) factor aggregation across four distinct market regimes: three recession-rich periods (1999–2022), a post-crisis recovery (2009–2022), a pre-crisis expansion (2003–2008), and a mid-cycle plateau (2013–2018). Three key findings emerge. <i>First, forecast horizon matters.</i> PC-based models dominate at the one-month horizon, reducing benchmark AR (1) mean-squared forecast errors (MSFE) by 10–45%, while BA delivers the lowest errors at horizons of three to twelve months. Cluster-based averaging surpasses BA only in exceptionally low-volatility conditions. <i>Second, backward-looking weighting adds limited incremental value.</i> The discounted-MSFE (DMSFE) scheme closely tracks BA’s performance but rarely outperforms it once bootstrap variance reduction is applied. <i>Third, a hybrid approach offers robust performance across cycles.</i> A two-step strategy deploying PC forecasts for 1–3&#xa0;months, switching to BA beyond three months, and activating cluster weights only for 6–12-month forecasts in stable regimes reduces MSFE by up to 24% in recessions and 12% in expansions. These findings suggest that practitioners should adapt their ensemble strategy to both forecast horizon and market regime: use PC factors for tactical timing, BA for medium-term views, and cluster averaging only under calm conditions. This adaptive approach maximizes the informational value of macro and technical signals while minimizing model risk across the business cycle.</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

Equity risk premium forecasting: a comparative analysis of bootstrap bagging and combination forecast

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Haithem Awijen,
  • Hachmi Ben Ameur,
  • Zied Ftiti

摘要

This paper evaluates whether forecast ensembles improve out-of-sample predictions of the monthly U.S. equity risk premium, relative to standard single-predictor ARDL regressions. Using thirty macro-financial and technical indicators, we assess the performance of boot-strap bagging (BA), five linear combination schemes, and principal component (PC) factor aggregation across four distinct market regimes: three recession-rich periods (1999–2022), a post-crisis recovery (2009–2022), a pre-crisis expansion (2003–2008), and a mid-cycle plateau (2013–2018). Three key findings emerge. First, forecast horizon matters. PC-based models dominate at the one-month horizon, reducing benchmark AR (1) mean-squared forecast errors (MSFE) by 10–45%, while BA delivers the lowest errors at horizons of three to twelve months. Cluster-based averaging surpasses BA only in exceptionally low-volatility conditions. Second, backward-looking weighting adds limited incremental value. The discounted-MSFE (DMSFE) scheme closely tracks BA’s performance but rarely outperforms it once bootstrap variance reduction is applied. Third, a hybrid approach offers robust performance across cycles. A two-step strategy deploying PC forecasts for 1–3 months, switching to BA beyond three months, and activating cluster weights only for 6–12-month forecasts in stable regimes reduces MSFE by up to 24% in recessions and 12% in expansions. These findings suggest that practitioners should adapt their ensemble strategy to both forecast horizon and market regime: use PC factors for tactical timing, BA for medium-term views, and cluster averaging only under calm conditions. This adaptive approach maximizes the informational value of macro and technical signals while minimizing model risk across the business cycle.