Abstract
This commentary provides my remarks on “Network Structure and Its Impact on Commodity Markets” by John Birge.
The message of Birge (2021) as I see it is that inverse optimization, and generalizations of this, can be used to uncover hidden supply network structures, enabling counterfactual analyses. These analyses can involve supply disruptions, policy interventions or similar changes in the environment to an industry or within its structure. The advantage of the inverse approach is that it takes into account the behavior of the participants in the industry, as well as the known facts about supply, demand and the network infrastructure. Assuming that the behavioral patterns of the market participants do not change fundamentally, for example, that their degree of risk aversion remains unchanged, such counterfactual analyses allow for projecting the industry's response to many kinds of changes in supply or demand. This is particularly useful in networked industries, where seemingly small changes in one part of the network can have large effects in other parts. It allows for policy makers to evaluate past policies and fine‐tune future changes to interventions. It allows for firms with major stakes across the network to gauge the impact of expansions, retirements, relocations, or upgrading infrastructure.
An alternative framework for such analyses includes various levels of bottom‐up modeling, where a different set of assumptions form the base premise. A typical approach in the analysis of electricity markets, for example, is to assume that there is a perfectly competitive market, or equivalently, a perfectly benevolent social planner, who is perfectly rational and with superb abilities regarding information gathering and forecasting (Skar et al. (2014)). This information gathering involves knowing all present and future demand and supply details including costs and capacity parameters. In sophisticated analyses, the future is viewed as exhibiting risk, with specified probability distributions or stochastic processes for important uncertainty factors. This framework can shed light on what might be a rational way forward, given full information and razor‐sharp optimizing behavior on the part of operators, investors, and consumers.
Another alternative may include agent‐based modeling, where demand and supply agents interact over a network (or with network agents), in a repeated game where the decision‐making structure of each agent must be tailored for each group. This setting allows agents to have and use limited information and exhibit bounded rationality, for example regarding the ability to anticipate future behavior of competitors. The ability of each agent type to optimize the path of consumption, operation, and investment must also be specified. Although pre‐specified types of learning can be built in, it is challenging to set the right information sets and optimizing behavior. So, the counterfactual analyses that come out of such an exercise must be judged in the light of the designed and often static bounded rationality of the agents constituting the market.
In contrast, Birge's (2021) examples use observed time series of decisions and prices as input to the inverse structural analyses. This ensures that at least how market participants have behaved in the past is captured. The way decision making has happened in the past is assumed to inform the analyst about relevant aspects of the basis for future decision making. By basis for decision making I mean, for example, parameters of cost/utility functions or parameters describing the network (Su and Judd (2012)).
The potential for application goes beyond risk propagation in partially observed networked industries. Examples include analysis of competitor behavior and reactions, and policy design. It opens new ways to analyze and understand supply chain phenomena.
Birge (2021) makes it seem easy. However, using inverse methods for counterfactual analyses in networks is challenging. There is always a lack of data. It is difficult to capture all relevant aspects of network decision making. Inverse problems come with their own challenges, as they are often “ill‐posed” and identifying core parameters is frequently difficult. These quantities include unobservable transaction costs (Fleten et al (2020)), as Birge mentions, but also ownership and contractual structures, which may be hard to uncover.
A useful element that is used in counterfactual analyses in the field of industrial organization is to assume that each decision observed is subject to randomness that can be accounted for by the decision makers but not by the analysts (Rust (1994)). This allows for better capturing the heterogeneity of decision makers, thus modeling differences in how they translate available information into concrete decisions. It also allows for the use of statistical methods for inferences about the unobservable network factors. Although not stated in Birge (2021), heterogeneity tends to lead to biases in uncovered structures (Gamba and Tesser (2009)).
Birge (2021) focuses on inverse optimization and three examples. It is possible to combine this methodology with simulation approaches, or even create hybrid techniques that integrate inverse optimization and bottom‐up procedures or agent‐based modeling.
In conclusion, I support the message of Birge (2021).
