Twenty-Eight Studies Later: AI Sharpens Freight Forecasts and Still Cannot See the Shocks

A systematic review of 28 machine-learning freight-forecasting studies published between 2012 and 2024, reported this week, reaches a conclusion the vendor deck will not: the models are genuinely better at the predictable part of the market, and structurally blind to the part that actually moves rates. What the research found: the literature shows growing and increasingly sophisticated use of vessel supply, commodity demand, bunker prices and macroeconomic indicators to predict markets, with neural-network architectures most common and hybrid or specialised models often outperforming standalone ones. But the review also concludes that geopolitical shocks, regulation and sudden supply disruptions cannot be captured reliably from historical data — which is a polite way of saying that a model trained on the past cannot price the Strait of Hormuz. The industry voices are unusually blunt. Roar Adland of SSY argues that large language models add little to the machine-learning systems already in use for more than a decade, and remain unsuitable for autonomous commercial decisions. Burak Cetinok of Arrow notes that modern tools process far larger datasets far faster, but weak data quality and limited coverage still constrain accuracy — the same binding constraint DNV named last week from an entirely different direction. The framework worth stealing: the experts identify three drivers of freight rates — predictable fundamentals, unpredictable shocks, and short-term herd behaviour. AI improves the first. It does nothing for the second. And it may make the third worse: if enough participants run similar models on similar data, they crowd into the same positions, accelerate the move and undermine the forecast that produced it. That is not a hypothetical failure mode; it is the documented history of quantitative finance. Why this matters for ship managers: this is the cleanest available answer to the question your board will eventually ask, which is whether an AI chartering tool can replace judgement on the desk. The evidence says it sharpens analysis, surfaces patterns and improves timing — and does not remove the cycle. It is decision support, and priced as decision support it is probably worth having. Priced as foresight, it is a way of paying to be confidently wrong at the same moment as everyone else. What to do now: when a forecasting vendor pitches you, ask two questions. Which of the three drivers does the model address? And how did it perform through the last genuine shock — not backtested, but live, in production, with a timestamp. If they cannot answer the second, you are buying the first driver only, and you should pay accordingly. The honest caveat: this is a literature review, not new primary research, and its window closes in 2024 — before the current generation of models. It is also reported secondhand through trade press rather than read here in full. The scepticism from Adland and Cetinok is genuine expert opinion, but both sit at brokers with a commercial stake in human judgement remaining valuable. [Splash247 · Maritime Economics & Logistics review]

This story is part of the Maritime AI Digest of 9 August 2026.