Scholarlink

<p>We contribute to the burgeoning literature on macroeconomic forecasting by exploring the benefits of machine learning (ML) models. We compare the performance of regularized regressions, tree-based models, and factor models in forecasting China’s real GDP growth rate through an extensive pseudo out-of-sample simulation, with forecast horizons ranging from 1 to 4 quarters ahead, over a period that stretches from January 1995 to December 2022. The forecasting performance is evaluated by using relative mean squared error, equal predictive test, and model confidence set, setting iterated AR as the benchmark. We find that ridge regression, tree-based methods, and FAAR-PCA substantially outperform the benchmark and other competing models in both the full sample period and the Covid period, achieving an averaged forecast variance error reduction of 73. 1% for the full sample period and 76.7% for the Covid period. The results indicate that ML algorithms are comparable to the golden standard factor models, enhancing China’s government’s ability to anticipate economic trends and tailor policy responses accordingly.</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

Understanding Machine Learning-based Methods in Macroeconomic Forecasting: Tracking Chinese GDP Growth Rate

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Qin Zhang,
  • Yong Bian

摘要

We contribute to the burgeoning literature on macroeconomic forecasting by exploring the benefits of machine learning (ML) models. We compare the performance of regularized regressions, tree-based models, and factor models in forecasting China’s real GDP growth rate through an extensive pseudo out-of-sample simulation, with forecast horizons ranging from 1 to 4 quarters ahead, over a period that stretches from January 1995 to December 2022. The forecasting performance is evaluated by using relative mean squared error, equal predictive test, and model confidence set, setting iterated AR as the benchmark. We find that ridge regression, tree-based methods, and FAAR-PCA substantially outperform the benchmark and other competing models in both the full sample period and the Covid period, achieving an averaged forecast variance error reduction of 73. 1% for the full sample period and 76.7% for the Covid period. The results indicate that ML algorithms are comparable to the golden standard factor models, enhancing China’s government’s ability to anticipate economic trends and tailor policy responses accordingly.