Mixed Frequency Deep Factor Asset Pricing with Multi-Source Heterogeneous Information on Policy Guidance

Abstract

In the era of big data, asset pricing depends on numerous factors from multi-source heterogeneous information, such as high frequency market and sentiment information, as well as low frequency firm characteristic and macroeconomic information. Actually, low frequency policy information plays a significant role in the long-term pricing but is rarely involved due to its textual form. To this end, we extract policy pricing factors from national strategies (“Five-Year Plans”, “Government Work Reports”, and “Monetary Policy Reports”) using the natural language processing (NLP) technique and dynamic topic model (DTM). Incorporating these low frequency policy factors to asset pricing will raise two issues: mixed frequency data and nonlinear effects. In this context, we propose a mixed frequency deep factor asset pricing model (MIDAS-DF) through introducing the mixed data sampling (MIDAS) technique and the deep learning (DL) architecture into the factor model structure. The MIDAS-DF model is able to learn nonlinear joint-patterns hidden in multi-source mixed frequency data, and output time-varying latent factors and factor loadings. We conduct an empirical analysis on 4939 stocks in the Chinese A-share market from January 2003 to July 2022. We find that the MIDAS-DF model outperforms several competing models in pricing on individual stocks, various test portfolios, and investment portfolios. The results from the MIDAS-DF model demonstrate that low frequency policy information imposes profound impacts on asset pricing and anchors the long-term pricing trend, while high frequency market and sentiment information optimize the short-term pricing accuracy. As a result, they work together to enhance the pricing performance.

Keywords: Asset pricing, Multi-source heterogeneous information, Policy guidance, Mixed data sampling (MIDAS), Deep learning, Latent factor model

JEL Classification: G12