A1942
Title: Deep learning with time contextual data
Authors: Eike-Christian Brinkop - University of Reading (United Kingdom) [presenting]
Emese Lazar - University of Reading (United Kingdom)
Marcel Prokopczuk - Leibniz University Hannover (Germany)
Abstract: Multiple analyses are conducted regarding the design of deep learning in asset pricing. A benchmark of objective functions is provided in a stock market setting and it is concluded that an economic foundation of the loss function benefits economic gains by resulting portfolios and control over the portfolio construction. Fundamental design optimisations are performed in an extensive hyperparameter pre-optimisation using the Hyperband algorithm for all the algorithms, maximizing their potential whilst completely removing model selection bias. In a pre-optimised hyperparameter setting, there is no economic benefit of performing dimension reduction techniques as a pre-processing unit in asset pricing or return prediction. The hypothesis is tested whether past information gives context for future asset returns. Using state-of-the-art machine learning algorithms from natural language processing and computer vision, the relevance of past information is analysed and the context they add to a pricing task. These structures are found to excel at building portfolios using contextual information of the underlying assets and compared to their feed-forward neural network counterparts without time contextual information. The filters provided by these two architectures are interpretable and help decode the black box that is machine learning in asset pricing. They allow gaining insight into the reaction of the market to companies' financial and market performance.