Supply chain momentum strategies with graph neural networks

Supply chain data has been studied as a potential source of investment opportunity for years. Many possible theoretical approaches can be employed when considering how an investor could leverage knowledge of supply chains. One of the most straightforward approaches, the customer momentum strategy, posits that if a company’s stock goes up, that is good news for that company’s suppliers — a kind of economic “as above, so below.” The key economic theory here is not so much that money flows back to supplier companies (this is too obvious to survive even the weakest form of an efficient markets hypothesis), but that investors in one company may be less aware of the financial state of companies linked to theirs via the supply chain. Those investors may also lack the expertise or mandate to invest in the related companies, thus creating an additional lead-lag effect.

Machine-learning techniques and the computational resources available to deploy them have advanced greatly since the initial research into supply chain data. Bloomberg researchers set out to investigate the use of one relatively new machine-learning technique, the Graph Neural Network (GNN), to construct an investment portfolio based on supply chain data. Bloomberg’s Supply Chain dataset is particularly well-suited to this because of the broad coverage, long history and additional quantified information available via our estimated relationship values.

Bloomberg’s supply chain data

Bloomberg’s Supply Chain dataset covers a global set of more than 20,000 companies and provides information on customer-supplier relationships between those companies and others. Since each company can have multiple relationships, the dataset contains over 100,000 companies in total. The overall dataset begins in 2006, with Bloomberg analyst estimates of relationship amounts being added to the offering starting in 2011. The estimated values give users of the data more information on quantified relationships than would otherwise be available.

Classical customer momentum strategies

The Bloomberg researchers investigated a number of non-machine-learning approaches as benchmarks:

• Classical momentum and reversal and strategies (without any supply chain data)
• A vanilla customer momentum strategy that ranks companies by the recent returns of their largest customer
• An improved classical strategy that aggregates the recent returns of a company’s customers by their pro rata contribution to the company’s revenue

For all of these strategies, they also tested several refinements, including varying the percentile threshold for what makes a long or short and only investing in companies with robustly covered supply chains. The best classical customer momentum strategy was able to achieve impressive results, with a high alpha even after attempting to attribute the returns to a Fama/French five-factor model.

Machine-learning strategies

Since any given company in the supply chain dataset may be connected to many other companies, it is natural to consider the entire supply chain dataset as one large graph. In such a graph, each company is a node, and its supply chain relationships are represented by directed edges. This perspective allows for the use of tools developed specifically for analysis of graphs, in particular, the use of a  graph neural network to generalize the analysis significantly:

• Adding more features of the customer companies by taking into account their market cap, volatility and turnover and considering different lookback horizons
• Incorporating additional supply chain relationship features to develop an optimal relationship weighting scheme
• Accounting for propagation effects from downstream customer firms

The graph neural network was trained with various parameters during testing. On the settings that worked best, the machine-learning model was able to generate an improved Sharpe Ratio compared with the classical strategies, and its alpha was still robust to a Fama/French five-factor attribution.

Download the white paper to learn more about how deploying a machine-learning strategy can improve the results of investing based on supply chain data.