Abstract:

Asset pricing explores the potential and latent factors that drive the prices of and returns on real and financial assets. In the big data era， asset pricing is influenced by many factors， such as high frequency market information， low frequency firm characteristic information and macroeconomic information. In this sense， typical mixed frequency data issue then arises. In addition， test assets play a critical role in the process of estimating pricing kernels. However， artificially constructing test assets to examine and optimize asset pricing models can not fully identify weaknesses in model pricing. How to efficiently capture useful pricing factors from multi-source heterogeneous information and achieve accurate asset pricing is valuable to practitioners， regulators， and academic researchers alike. To this end， the concept of mixed frequency stochastic discount factor （MF-SDF） is proposed through introducing a mixed data sampling （MIDAS） approach into the stochastic discount factor （SDF） theory. MF-SDF is used as a criterion function to construct a generative adversarial network （GAN） which solves the asset pricing issue under the mixed frequency data environment. Taken together， the proposed MIDAS-SDF-GAN model is able to handle mixed frequency data via the MIDAS approach and model the complex nonlinear pattern hidden in high-dimensional pricing factors through deep learning. It can employ the generative adversarial game mechanism to construct test assets， guiding itself to learn pricing patterns and continuously improve the pricing power in a data-driven way. All Chinese A-stock listed firms from January 2000 to December 2020 are chosen as our sample data and construct a large mixed frequency dataset containing 48 high frequency market variables， 153 low frequency firm characteristic variables and 11 low frequency macroeconomic variables. The data come from the CSMAR database. Then， the performance of the MIDAS-SDF-GAN model in asset pricing is empirically investigated. The results show that： 1） The MIDAS-SDF-GAN model is able to construct test assets in a data-driven way and guide itself to learn the pricing mechanism. It outperforms the other four competing models in both stock return prediction and investment profitability. 2） The MIDAS-SDF-GAN model is able to fully exploit information in mixed frequency data. It anchors the pricing basis through low frequency firm characteristic information and macroeconomic information， and optimizes the pricing accuracy through high frequency market information. These two types of information work together to improve the pricing effect. The study is an example of applying deep learning methods to the field of asset pricing. It is of great significance to solve the problems encountered in the process of asset pricing with the help of advanced artificial intelligence technology under the framework of financial theory.

Key words: asset pricing, stochastic discount factor, mixed data sampling （MIDAS）, generative adversarial network （GAN）, test assets