Scholarlink

<p>Forecasting stock returns in Asian equity markets is difficult because these markets exhibit stronger retail participation, more frequent policy intervention, and greater structural noise than mature developed markets. To address these features, we propose a Conditional Variational Recurrent Autoencoder (CVRA), a dynamic latent-factor model that combines recurrent neural networks with variational autoencoders. The recurrent component captures temporal dependence, while the variational structure models latent uncertainty and helps separate persistent risk from transitory noise. A characteristic-based beta network further maps lagged firm characteristics into time-varying factor loadings. Using monthly stock data from 2000 to 2024 in China and Japan, we show that CVRA outperforms traditional factor models, tree-based methods, and deep-learning benchmarks in both statistical and economic performance. Additional evidence indicates that the extracted latent factors are economically interpretable and closely related to macro-financial conditions. Overall, CVRA provides a flexible and economically grounded framework for asset pricing in Asian markets.</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

CVRA: Asset Pricing via the Conditional Variational Recurrent Autoencoder in Asian Market

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Yilun Wang

摘要

Forecasting stock returns in Asian equity markets is difficult because these markets exhibit stronger retail participation, more frequent policy intervention, and greater structural noise than mature developed markets. To address these features, we propose a Conditional Variational Recurrent Autoencoder (CVRA), a dynamic latent-factor model that combines recurrent neural networks with variational autoencoders. The recurrent component captures temporal dependence, while the variational structure models latent uncertainty and helps separate persistent risk from transitory noise. A characteristic-based beta network further maps lagged firm characteristics into time-varying factor loadings. Using monthly stock data from 2000 to 2024 in China and Japan, we show that CVRA outperforms traditional factor models, tree-based methods, and deep-learning benchmarks in both statistical and economic performance. Additional evidence indicates that the extracted latent factors are economically interpretable and closely related to macro-financial conditions. Overall, CVRA provides a flexible and economically grounded framework for asset pricing in Asian markets.