Applying AI to decision-making in shipping and commodities markets
In this installment of the AI in Supply Chain series (#AIinSupplyChain), we explore the topic of decision-making in the shipping and commodities markets.
Before we proceed, it is important to note four characteristics of the freight shipping industry that were highlighted by Roar Adland, a professor of shipping economics at the Norwegian School of Economics. In an August 2017 blog post on LinkedIn: 4 things shipping had long before Uber, he noted the following: First, shipping inherently utilizes dynamic pricing because of the volatile nature of rates, and this has been the case for a few centuries. Second, the industry already matches demand and supply in a highly efficient manner. Third, the industry fully utilizes its assets — when there are empty ships it is because there is no demand, not because demand and supply are being poorly matched. Fourth, shipowners have elevated regulatory and tax arbitrage to a fine art, and they have been doing this since the 1970s.
Adland’s observations helped to confirm conclusions I had reached leading up to and following the publication of Industry Study: Ocean Freight Shipping (#Startups) in February 2017 and Updates — Industry Study: Ocean Freight Shipping (#Startups) in June 2017. There’s more than meets the eye when it comes to innovation in shipping, and the “Uber for shipping” business model lacks enough nuance to succeed in the manner startup founders and most venture capitalists assume it should.
In this article, I will briefly highlight three examples of how machine learning and artificial intelligence are being applied to problems in dry bulk shipping.
Making better decisions with data science
Paul Doetsch is the CEO of True Bearing Insights, a Boston-based startup that uses the power of data science and predictive analytics to aid commercial decision-making in commodities and bulk shipping markets. I have known him since March 2017, when we met at the Connecticut Maritime Association’s Shipping Conference and Exhibition. I spoke with him in July about what True Bearing Insights does.
True Bearing Insights is a decision-support platform that uses computers and software to do what machines do best — crunch through vast troves of data while looking for patterns, while simultaneously enabling people to do what humans do best in the decision-making loop — manage exceptions and other unfamiliar situations. True Bearing does not think machine learning will replace human judgment, but rather that machine learning will make people more effective when human judgment matters most.
I asked Doetsch to explain True Bearing’s approach a bit more. “As a defining characteristic of any business, we all want to make better decisions,” he said. “Understanding both the question, and the context, forces us to step back and look at the overall decision-making process to identify constraints. Do we have the right data or information to make better decisions? Do we even know what the right data is? Are we drowning in data? At some point, a human’s capacity to systematically manage information becomes the constraint.”
At this point, True Bearing Insights is using freight forward agreements to demonstrate the power of its platform, which became operational early this year. So the software can help shipowners make decisions about how to trade freight derivatives on iron ore, coal and other dry bulk commodities, but it can also be used to answer questions like, “Should I run Australia or Brazil?”
According to Doetsch, this matters more than people outside the industry realize, because the commodities transported by the industry are produced primarily in the Southern Hemisphere while demand for those commodities is overwhelmingly located in the Northern Hemisphere.
As a result, Doetsch argues that a focus on asking the right question is critical and further that different questions must be asked at different points in time because so much about the future is unknown.
He says, “Is the question we are asking designed to provide another data point for a constrained human decision-maker to be better informed, or does it go further downstream to make specific recommendations to supplement human judgment?”
During our conversation, he emphasized that clarity and specificity around the questions being asked are requisite to success, arguing that clarity and specificity make it possible to define objective measures of success in the context of the overall business and to determine how this impacts the income statement, the balance sheet and the cash flow statement.
Observing that technology like that being developed by True Bearing Insights makes it possible for shipowners to ask and answer more nuanced questions than they could in the past, he said, “Understanding the question uncovers another layer of complexity — aligning performance measures of a new system with actual value created. The question being asked is usually more complex than initially apparent, and usually must be broken down into a sequence of steps and models or representations of the real world.”
Finally, he concludes, “Digitization is simply a means to an end — it is the clarity around the question itself that catalyzes truly transformative solutions.”
Machine learning as decision-support system
Herman Bomholt and Torsten Thune made the news when their master’s thesis demonstrated that certain machine learning algorithms outperformed the relevant benchmark indices by about 10% or $1,700 per day between 2017 and 2019. I spoke with them earlier this month about their research.
The thesis is undergoing peer review, and so full details have not yet been made public. However, we discussed the implications of their work and what the implications are for the shipping industry more broadly. Below, I paraphrase and summarize our conversation.
First, there’s an important difference between what is theoretically possible and real-world constraints. This is especially true when it comes to obtaining data that is of the requisite quality to power machine learning systems. Particularly, people with domain expertise are critical to the formulation of the assumptions on which machine learning algorithms depend. That is why machine learning and optimization should be treated as decision-support systems rather than as replacements for human expertise. The formula we should be aiming for is human PLUS machine, not human OR machine.
Second, market cycles in the shipping industry matter and make the task of developing and maintaining machine learning and optimization systems less straightforward than one might expect. Some of the issues that must be addressed include figuring out what the appropriate time spans are for the investigations being made; the frequency of available data and if that matches requirements; whether there is a long enough history of data about the phenomena being studied; accounting for the fact that shipping routes are constantly changing; and accounting for changing trip charter rates. However, Bomholt and Thune believe that individual shipping companies probably keep more detained and complete data than they could easily find from public sources of industry data.
I asked Bomholt and Thune for a few actionable suggestions they believe people in shipping who are exploring using machine learning and AI in running their business operations should consider.
First, they suggest that shipping companies invest more in gathering proprietary and nonproprietary data and to use that data in addressing problems the companies are facing.
Second, they suggest that analytics teams should comprise people who understand machine learning really well AND people who have an intimate understanding of shipping industry data and analytics. They suggest these teams should start slowly and test their machine learning algorithms against past data and events before using them to forecast future decisions. Even then, they emphasize that human judgment must be part of the decision-making process.
Third, they suggest making assessments about the following questions: What does the machine understand? What can the machine not understand? And what needs further research by people?
The insights Bomholt and Thune shared about the need for human-machine decision-making systems in shipping agree with observations by David Sidoti in his 2018 Ph.D. dissertation Novel Optimization-based Algorithms for Dynamic Resource Management in Complex Systems and in a 2015 paper, Dynamic Resource Management and Information Integration for Proactive Decision Support and Planning, authored by Sidoti and others and presented at the 18th International Conference on Information Fusion in Washington in 2015.
The value of foresight in dry bulk market
In January 2018, Adland, Vit Proachazka and Stein W. Wallace published The value of foresight in the drybulk freight market. The authors considered “an optimization problem that all operators of vessels in the bulk shipping freight markets face: How to optimally reallocate a ship or fleet of ships through space and time by sequentially accepting freight contracts (charters) for spot market cargoes between port pairs in a transport network, often called tramp shipping.”
They solved two versions of the problem. First, they solved the case in which they assume that future rates are known with perfect foresight. Second, they solved the case in which they assume that future rates are known with limited foresight.
In conclusion the authors find that, first, with perfect knowledge, their model yields higher cumulative earrings between 2006 and 2016 — though they point out that the financial crisis creates a distortion in freight rates up to late 2008. This is obviously unrealistic but is important for establishing a theoretical maximum — which they found to be 15% between 2006 and 2016, but was higher between 2009 and 2016, at 20%, 15% and 23% for the Supramax, Panamax and Capesize sectors, respectively.
Second, when they assumed perfect knowledge on a limited and relatively short time horizon, they found that it is possible to capture a large portion of the theoretical maximum that they established by assuming perfect foresight. They point out that this required a different methodology, and that the assumption of perfect foresight on a limited horizon is still unrealistic.
Third, they found that “it is also possible to use the optimization model to provide further insight into the market. For instance, we have calculated the value of each decision at every point in space and time. With that, we have observed an asymmetry in the geographical switching function for the Atlantic region. That is, an incorrect decision “go to Pacific” is potentially more costly if applied at the wrong time than an incorrect decision to “stay in the Atlantic.”
Overall they conclude that “the empirical findings reveal a big potential of exploiting spatial inefficiencies by a sophisticated chartering strategy. A natural continuation of this research would be to apply stochastic programming to handle uncertainty in freight rates. For instance, to use a scenario tree for describing the future development of freight rates instead of the assumption of perfect foresight.”
Given the sheer size, scale and importance of shipping as a driver of global commerce across all industries, early-stage technology investors are fascinated by the notion that digital technologies will disrupt business processes and norms in the shipping industry. Based on the preceding discussion, readers of FreightWaves who are not more familiar with the industry would be forgiven for thinking that we are on the cusp of an explosion of machine learning and AI applied to dry-bulk shipping and commodities markets in the near term — say the next two to five years.
Overall, I am highly skeptical.
Between November 2017 and October 2018, I served as a volunteer advisory board member of the New York Maritime Innovation Center. My experience over that period helped me realize that innovation in the maritime industry is a slow process that depends on top-down mandates which are complicated by complex and interlocking regulatory regimes and compliance requirements.
This does not mean innovation will never happen. However it means that things will move much more slowly perhaps even than some veteran industry participants would prefer.
That said, I expect that academic researchers like Adland and his colleagues at the Norwegian School of Economics, and early-stage technology startup founders like Doetsch and his teammates at True Bearing Insights will continue pushing the envelope while they wait for the rest of the industry to catch up with them.
This also means that executive teams in the industry that are willing to push the envelope and experiment with machine learning, AI and other predictive analytics technologies, relatively speaking, could establish advantages that endure longer than one would otherwise assume.
While preparing to write this article, I emailed a C-level executive at a large container shipping company to ask about applications of AI in the industry. He responded, “AI in Container Shipping is, in my opinion, still a bit far fetched and there are so many pressing issues to tackle that these are not at the top of my priority list at this point. … Obviously have some developers working in the BI (business intelligence) section that fumble with AI but I would not venture to call them experts as yet.” He however pointed out that this is not his area of expertise.
Source: Freight Waves