Editor’s Note: In the current world of new media and new analytical tools such as AI (artificial intelligence), some might try to use them to evaluate the strengths and weaknesses of complex railroad mergers. Two industry veterans offer this “alert message” on possible challenges with complex enterprise merger evaluation techniques: Contributing Editor Jim Blaze with his USRA and Conrail experience, and Chris Rooney with his policy and financial experience. Both are graduates of the University of Chicago, with different skill sets. They represented opposing parties during the “Let Conrail be Conrail” era. Here, they comment on “theoretical dependence upon early-generation AI-based tool kits” and their “risk/reward-profiles.” – William C. Vantuono
Now afoot: an $85 billion dollar megamerger being pursued by Union Pacific and Norfolk Southern. How is that prospective outcome going to be commercially and geopolitically examined as the companies seek regulatory approval, while others oppose it?
What if we used an AI model to analyze the UP+NS merger, for example, OpenAI-o3, Anthropic Claude Sonnet, Perplexity or Llama? What can we two veteran merger hands infer from using these new logic tools?
From a top-down level, UP+NS looks like a marriage made in heaven Top table), with high stock valuation and equity money needed to put together a merger of this scale. This is particularly so if you make the “collateral assumption” that BNSF must then buy CSX. But is that a necessarily true condition? The bottom table below isolates the few strategic resulting enterprise key sizing metrics of the two theoretical resulting rail systems. UP+NS vs. competition with a follow-on BNSF+CSX merger. At an elevated level, we, as long-term spreadsheet- and database-using business analysts with hands-on experience, do see well-balanced aggregate metrics between the top two likely to emerge east-west expanded rail networks.
If we substituted AI tools, what’s the learning curve? Are they necessarily better? Are the initial AI tools ready, capable of detailed insight? Or are they “inferring” at a less precise commercial and operational level?
“Old school” pre/post-merger analysts would focus on network and pro forma models developed between 1970 and 2000 for an assessment of realistic operations/marketing and financial outcomes. That is not how analysts using new AI systems might initially proceed with their risk/reward assessment. The AI technician will of course focus upon data sets. But will initial railway freight sector data contain biases? We can’t be sure yet.
Example: So called “Grok Techniques” would harvest a vast amount of raw data, including all relevant entries and then add something from ChatGPT. But how rich is a place where railroad analysis and critical data sets are possibly found? Do we know?
The early use process for employing AI might very well perpetuate some critical biases, leading to unfair or inaccurate risk assessments. Is that not a prudent risk expectation? For example, an AI tool used for loan applications might unfairly deny loans to certain demographics if the AI model training data reflects historical lending biases.
Furthermore, many AI models, especially deep learning models, are really “black boxes,” meaning they are difficult to understand and explain in order to produce an audit of how they arrive at their decisions. The black box developers might even try to disguise from STB regulators how their “private box” works. Any lack of transparency can make it challenging to identify the root cause of errors or biases in the risk assessment process. It can also make it difficult to comply with existing and still important safety and pro-competitive regulations that require explainable decision-making, as in a court-like public review.
Over-reliance on AI for risk assessment can also potentially result in a missed interpretation of how internal corporate culture and human behavior can disrupt the projection of complex merger execution events. In contrast, skilled people with technical case study and audit assessment experience can often identify subtle nuances and contextual factors that AI might miss, due to its early formatting as a tool of equations and algorithms.
Where and how does such human “touch” come into play in a complex, high-dollar risk/reward assessment process that uses early versions of AI logic models? AI, by broad admission, is in the early stages of equation and algorithm testings. “Be cautious” is our message. Over-reliance on AI for risk assessment can lead to a lack of experienced human insight. It all depends on your point of view, whether you are an owner or user.
ECONOMICS OF NETWORK GEOGRAPHY
The geography of the deal presents another level of data and yet another level of operational subtleties. Railroads are the quintessential physical undertakings, operating on the ground, not in the cloud, on fixed rails well within the laws of physics—more so even than trucks. No tracks, no service, unless deals can be cut with whoever owns the rails.
Test, but expect some errors. A few critical errors may change the resulting outcomes, perhaps negating old predictions of market share shifts among freight modes, and upending otherwise-probable sector volume expectations by corridor or other segmented definitions.
Let’s take a geographical look at the UP+NS and then BNSF+CSX by studying the coverage of the rail tracks themselves.
The above maps suggest possible weaknesses in both the proposed UP+NS merger and the hypothetical combination of BNSF and CSX to the trained eye. Areas that appear underserved bythe UP+NS deal are virtually all New England (NS has access to Ayer in central Massachusetts) and most of Florida beyond Jacksonville and anything north of Minneapolis.
Likewise, BNSF+CSX misses a large part of the Central Corridor (BNSF has the limited right to use UP tracks from Denver to Northern California) and its access to Mexico would be more precarious if it could not reach the CPKC in Mexico (via UP) and had to rely solely on its connections with the other Mexican carrier, Ferromex (26% owned by UP) at Eagle Pass and El Paso.
Drilling down farther, all tracks are not created equal.
That all tracks do not receive equal traffic volume is illustrated by the above graphic, which is the result of the Federal Railroad Administration flowing the STB mandated statistical sampling of individual freight invoices (waybills) across the rail system with tons as the common denominator.
Thus, “to and from specifically where?” becomes a second-order concern not encountered with other modes such as air and trucking. For example, take Florida: Setting aside the West Coast, which would require a deal with CSX to reach Tampa, how well is the East Coast served? Yes, NS could reach the East Coast on existing lines to Jacksonville and, yes, it could hand off traffic to Florida East Coast, as it does today.
But one level lower, as a third-order concern, are routing decisions that solve access problems but may increase costs. From New Orleans for example, the options for routing traffic over NS up to Birmingham and then to Atlanta and back down produce 20% more miles than today’s usual routing via Mobile on CSX.
And, as a fourth order concern, many of the freight cars carrying carload traffic such as hoppers and tank cars are rented to shippers on a mileage basis, thus complicating the calculation of who benefits, who must invest capital to succeed, and what factors or secondary decision makers (investors) have a significant role to play out in order to achieve the promised gains.
To mitigate these analytical weaknesses, the ideal oversight and commercial plus regulatory process would suggest the following roles and testing to improve AI’s evaluation contribution. Following is a simple checklist offered for discussion.
- Build robustness into the modeling of operations and critical data sets. Invest in data cleaning and bias detection and mitigation techniques to ensure objectivity. Specifically, rail mergers normally proceed at the STB with parties having reasonable access to the Rail Waybill Sample, a statistical sampling of all orders to move freight cars or containers from one point to another. This allows participants to study the probable effects of the merger on established freight flows.
- Prioritize explainability. Choose models that are inherently explainable or develop methods to make complex models more transparent.
- Implement test routines on version 1.0 toward 3.0 logic models. Protecting against being seduced by the early output versions is the message here. It might be too easy to succumb to such low-labor-cost “budget versions” without vetting.
- Maintain human oversight. Don’t rely solely on AI models. Ensure humans participate in the risk assessment process and that they can override AI formula structure decisions when necessary.
- Continuously search for the critical relevant data patterns and assess their impact on AI models’ conclusions if such models are used as conclusive evidence.
The prospects of these two mega railroad mergers are worth exploring. New tools are to be welcomed. But we should expect twists and turns in how predictable some expectations might or might not be. Certainly, achieving elevated levels of decimal place position in anyone’s pro forma impact model is tentative, way too early to confidentially call out.
Keep a prudent “discount rate” in your toolbox to be periodically reassessed over the next two or more years.
Independent railway economist and Railway Age Contributing Editor Jim Blaze has been in the railroad industry for close to 50 years. Trained in logistics, he served seven years with the Illinois DOT as a Chicago long-range freight planner and almost two years with the USRA technical staff in Washington, D.C. Jim then spent 21 years with Conrail in cross-functional strategic roles from branch line economics to mergers, IT, logistics, and corporate change. He followed this with 20 years of international consulting at rail engineering firm Zeta-Tech Associated. Jim is a Magna Cum Laude Graduate of St Anselm’s College with a master’s degree from the University of Chicago. Married with six children, he lives outside of Philadelphia. “This column reflects my continued passion for the future of railroading as a competitive industry,” says Jim. “Only by occasionally challenging our institutions can we probe for better quality and performance. My opinions are my own, independent of Railway Age. As always, contrary business opinions are welcome.”
Chris Rooney has more than 30 years of experience in transportation and financial management. As Deputy Federal Railway Administrator, he was a government spokesman before Congress and other government bodies and authored government positions on rail transport regulation and coordination. He led the federal government technical support team for the sale of U.S. government-owned Conrail to the private sector, eventually as an IPO. His work has included recommendations on capital and operating funding for railway assistance and on appropriate regulation. He has advised many railroads, railroad clients and federal, state and local transportation agencies on strategic planning, economic, contractual and financial issues in the U.S., Canada, Europe, Latin America, Asia and Africa. He has appeared as an expert witness in several ICC and Surface Transportation Board merger proceedings and was a member of the team re-examining the role of Constrained Market Pricing and the Stand Alone Cost principles used by the STB for rate regulation in quasi-monopoly situations. Rooney holds a Chartered Financial Analystprofessionaldesignation from the Institute of Chartered Financial Analysts.
