Preprints and early-stage research may not have been peer reviewed yet.
It is impossible to analyze an asset taken in isolation, without taking into account the wider picture of the market. This fact is behind the extensive use of copulas or vector autoregressive models in finance, which allow to model dependencies between assets. In this paper, we look at the graph-based method to model inter-asset behavior. Graphs are ubiquitous when representing relationships, whether to model social network interactions, disease spread, traffic, or supply chain information. It allows for a very intuitive representation of market interconnections. We show how several types of market information can be translated into graphs and show some graph-based tools for market analysis. Furthermore, neural convolution layers have been developed which allow for more expressive neural models. Just like Euclidean convolution layers on images, they promise to contextualise each individual asset during prediction. We show the effect of three graph neural layers on the stock return forecasting problem. Using these forecasts, we build portfolios and show that graph layers act as a stabilizer to classical methods like LSTM, reducing transaction costs and filtering out high-frequency signals. We also study the effect of different graph-based information on the forecast and observe that in 2021, supply chain information becomes much more informative than sectoral or correlation-based graphs.