Conditional autoencoder asset pricing models for the Korean stock market

Area of Science:

• Quantitative Finance
• Machine Learning
• Econometrics

Background:

• Traditional asset pricing models struggle to capture complex market dynamics.
• The need for advanced methods to analyze latent factors in financial markets is growing.
• Machine learning offers novel approaches to financial modeling.

Purpose of the Study:

• To evaluate the explanatory power of a conditional autoencoder (CA) model in the Korean stock market.
• To compare the CA model's performance against traditional asset pricing models.
• To investigate the role of firm characteristics under varying macro-financial conditions.

Main Methods:

• Utilizing a conditional autoencoder (CA), a machine learning technique, to extract latent factors.
• Estimating factor exposure as a flexible nonlinear function of covariates.
• Analyzing pricing errors to assess investment strategy performance.

Main Results:

• The CA model exhibits excellent explanatory power across the entire sample and subsamples in the Korean market.
• The CA model significantly improves the explanation of market anomalies compared to traditional models.
• Investment strategies based on the CA model's pricing error yield better expected stock returns.
• Firm characteristics are identified as crucial for asset pricing, conditional on macro-financial states like crises.

Conclusions:

• The CA model provides a superior framework for asset pricing in the Korean stock market.
• The model's ability to adapt to changing macro-financial conditions enhances its practical applicability.
• Future research can leverage the CA model for a more comprehensive understanding of asset pricing dynamics.