Investment portfolio built with graph neural networks demonstrates promising results
By Ather Fawaz · Aug 19, 2020
By Ather Fawaz · Aug 19, 2020
[SHOWTOGROUPS=4,20,22]
Для просмотра ссылки Войдиили Зарегистрируйся
Graph neural networks (GNNs) are a relatively recent development in the field of machine learning. Like traditional graphs, a core principle of GNNs is that they model the dependencies and relationships between nodes. A particularly nifty feature in this architecture is that unlike standard neural networks, Для просмотра ссылки Войдиили Зарегистрируйся that can represent information from its neighborhood with arbitrary depth.
Building on this, Bloomberg recently Для просмотра ссылки Войдиили Зарегистрируйся that leverages graph neural networks to construct an investment portfolio based on supply chain data. The researchers used Bloomberg's Supply Chain dataset, which contains data on over 100,000 companies, to develop an extension of the classical customer momentum strategy that was Для просмотра ссылки Войди или Зарегистрируйсяback in 2008.
Since a given company in the dataset may be connected to multiple companies, we can consider the entire supply chain dataset as one gargantuan graph. In this graph, each company would be represented by a node while its supply chain relationships would be represented by directed edges between the nodes. Bloomberg took this idea and mapped it into a problem that could be solved via GNNs. And the intrinsic architecture of GNNs allowed the traditional analysis of supply chain data to be generalized in multiple ways.
First, it allowed the inclusion of more features of customer companies by taking into account their market cap, volatility, turnover, and considering different lookback horizons. Similarly, GNNs were able to factor the propagation effects from downstream customer firms and also incorporated additional supply chain relationship features to develop an optimal weighting scheme.
The end results showed graph neural network model generated a long-short portfolio that demonstrated "an improved Для просмотра ссылки Войдиили Зарегистрируйся compared with the classical strategies and its alpha was still robust to aДля просмотра ссылки Войди или Зарегистрируйся attribution." This indicates the "usefulness of using a company's customer relationships from the supply chain dataset in conjunction with graph neural networks," Bloomberg wrote.
Moving forward, the researchers believe that by modifying the underlying graph structure to support different edge types, the model can be extended to include more features like the suppliers of a company and the features of a target company. For companies that already have a traditional supply chain model in place, a GNN can be used to test whether the graph structure of the supply chain dataset might increase the model's predictive power.
[/SHOWTOGROUPS]
Для просмотра ссылки Войди
Graph neural networks (GNNs) are a relatively recent development in the field of machine learning. Like traditional graphs, a core principle of GNNs is that they model the dependencies and relationships between nodes. A particularly nifty feature in this architecture is that unlike standard neural networks, Для просмотра ссылки Войди
Building on this, Bloomberg recently Для просмотра ссылки Войди
Since a given company in the dataset may be connected to multiple companies, we can consider the entire supply chain dataset as one gargantuan graph. In this graph, each company would be represented by a node while its supply chain relationships would be represented by directed edges between the nodes. Bloomberg took this idea and mapped it into a problem that could be solved via GNNs. And the intrinsic architecture of GNNs allowed the traditional analysis of supply chain data to be generalized in multiple ways.
First, it allowed the inclusion of more features of customer companies by taking into account their market cap, volatility, turnover, and considering different lookback horizons. Similarly, GNNs were able to factor the propagation effects from downstream customer firms and also incorporated additional supply chain relationship features to develop an optimal weighting scheme.
The end results showed graph neural network model generated a long-short portfolio that demonstrated "an improved Для просмотра ссылки Войди
Moving forward, the researchers believe that by modifying the underlying graph structure to support different edge types, the model can be extended to include more features like the suppliers of a company and the features of a target company. For companies that already have a traditional supply chain model in place, a GNN can be used to test whether the graph structure of the supply chain dataset might increase the model's predictive power.
[/SHOWTOGROUPS]
