Efficient at What?

On the morning of February 28th, while following the news of the first US and Israeli strikes on Iran, we were running a signal monitor.

Not because we were looking for a trade. We’d been watching European gas storage fall below seasonal norms for weeks, knowing that Hormuz had always been the obvious chokepoint for the LNG that Europe depends on to get through the heating season. If something happened there, the people who would feel it first wouldn’t be commodity traders. They’d be families opening their energy bills in October.

The monitor was a Python script, open-source data feeds, a handful of API calls. Storage levels from the GIE transparency portal. Tanker stocks on Yahoo Finance as a Hormuz shipping proxy. ISW situation reports. When the morning’s numbers came in, two signals fired simultaneously: gas storage still withdrawing deep into the heating season with the injection window not yet open, and tanker volumes in free fall.

The market agreed quickly. Brent crude rose from around $70 to above $84 within hours. TTF natural gas futures moved more aggressively still. Goldman Sachs, which had been forecasting €36/MWh for April, revised to €55 within the week. According to the Economics Observatory, prices had “almost doubled” from pre-war levels within days.

The event was recognized. The mechanism worked. The market knew.

“It priced it immediately,” Àngel said.

Yes. And then we both sat with what that actually means. Because the price moving doesn’t settle the question we were really asking. The market knowing that a war started is not the same as the market correctly estimating how long a disruption lasts, or how much it compounds before it resolves. Those are different capabilities. And the gap between them is where this story gets interesting.

What the public forecast may be getting wrong

Martin Wolf wrote in the FT this week that the oil price rise “matches that immediately after Russia’s full-scale invasion of Ukraine, but the gas price rise is far smaller.” He’s right about the data. The implicit conclusion, that this energy shock is less severe than the Ukrainian one, relies on a historical comparison that may not hold for long.

The published forecast consensus reflects the historical prior, updated for confirmed facts visible to everyone. What it appears to be underweighting is a cluster of mechanisms that don’t appear in any previous Hormuz analysis, because they never needed to.

Insurance. P&I war-risk coverage for vessels in the Persian Gulf was suspended on March 5th. In every previous Hormuz crisis, insurance markets remained functional. The geometry of risk was managed at the cargo level. This time, the suspension made shipping volumes a function of insurance market reset timelines, not just ceasefire diplomacy. Those timelines are measured in weeks to months. That variable doesn’t appear in previous Hormuz analyses because it never moved at scale.

Qatar’s structure. Qatar has no bypass pipeline. Ras Laffan offline is not a disruption that can be partially compensated through rerouting: it is a binary stop. Every week closed is a week of European storage that cannot be recovered retroactively. A ceasefire stops the damage from compounding. It doesn’t undo what’s already accumulated. Wolf’s framework accounts for supply disruption, but the asymmetry between closing and reopening a binary LNG terminal is not in any previous Ukraine-era model.

Asymmetric restoration. Fields that shut in, terminals that need damage assessment, reservoir pressure that takes time to restabilize. Ceasefire and gas flowing are different events on different timelines. The assumption that ending active hostilities restores supply quickly is plausible for pipeline gas. For LNG under suspended insurance, it isn’t.

These mechanisms are visible in public data. They’re not secret. They’re in shipping reports, insurance market notices, and infrastructure analyses. But they require something specific to see: not just access to the information, but enough accumulated context to know which prior to distrust.

Capital Economics put numbers to this in a scenario analysis published this week. In their most severe case (three months of conflict with lasting damage to Kharg Island), they project a loss of 8-9% of global oil and LNG exports, with repercussions through 2027, oil at $150 per barrel, EU gas at €120 per megawatt-hour. Their only comparable historical reference: the energy crises of the late 1970s to mid-1980s. That last detail is the one that matters most here. The prior that most public models are working from ends well before those years.

What changed on our end

We didn’t have anything the FT doesn’t have. The GIE portal is public. Kpler publishes shipping data. The P&I insurance suspension was reported by Reuters. ISW publishes daily situation reports for free.

What we had was something that retained context across weeks. When the insurance suspension happened on March 5th, we had been watching tanker stocks as a Hormuz proxy since early February. The connection between “insurance suspended” and “the variable that was never in the model” didn’t require a new data source. It required remembering which thread you were following.

The CFA Institute described something like this in 2025 as the collapse of processing asymmetry: the capacity to synthesize across many threads simultaneously, something previously reserved for teams of analysts, becoming available to individuals with access to AI tools. That’s what we had. Not better information. A different synthesis capacity. We could hold the gas storage thread, the insurance thread, the Qatar infrastructure thread, and the temperature forecast thread in parallel, across weeks, and notice when a new data point added to one of them.

For now, this requires wiring something together manually. A Python script. A few API keys. Weeks of context. Knowing which questions to ask.

That is about to change.

The paradox of democratization

The applications that will make this kind of synthesis accessible to anyone are being built. Not a script: something you open on your phone. Multiple AI assistants running in parallel, each synthesizing from the same public sources, delivering a research-quality briefing to anyone willing to read it.

When that happens, the math changes in a way the democratization story doesn’t fully account for.

Two people using different AI tools, reading the same GIE data, the same ISW reports, the same Reuters coverage, will reach similar theses. Not because either AI is wrong, but because models trained on overlapping corpora, starting from identical public inputs, will converge. The processing asymmetry between an individual and a specialist analyst may collapse. The epistemic asymmetry between the retail layer and the institutional layer will not.

What you get is a retail layer that is faster and better-informed about public data than it has ever been, and simultaneously more homogeneous in its conclusions. More perfectly imperfect, in a specific way: the signal is cleaner at the center of the distribution. Everyone correctly identifies the war. Everyone updates for the confirmed data. The tail events, the variables that have never moved before, the mechanisms that don’t appear in any training corpus or previous analysis, those still get missed. And they get missed by everyone at once.

The insurance market is exactly that kind of variable. Not a secret. Available to anyone who was following the thread. But it required knowing to look for it, because this time it mattered in a way it had never mattered before. A model trained on the history of Hormuz crises, updated by the public data available on March 5th, would not automatically weight maritime insurance as a key variable. Neither would most analysts anchoring to the Ukraine comparison.

The paradox is this: more AI research capacity at the retail level does not mean more diverse market views. It may mean faster and more coordinated convergence on whatever the public data supports, with greater collective blindness to the variables that fall outside the historical record.

The farmer and the thermometer

“You know what I keep thinking about?” Àngel said one morning, after we’d been working through the insurance mechanism. “A farmer who buys extra propane in October because this winter feels different. Not because of a model. Because he’s spent enough years watching the same sky to feel when something is breaking.”

He wasn’t romanticizing folk wisdom. He was pointing at something structural. Knowing to look at the insurance market wasn’t a processing capacity. It wasn’t something that would emerge automatically from running more API calls or adding more data sources. It came from following a causal thread: war at the chokepoint, shipping risk, insurance suspension, no sovereign backstop, ships don’t move, supply disruption is not just about missiles or diplomacy.

What AI gives you is the ability to hold that thread across many sources simultaneously, to retain it across weeks, to notice when a new piece of information adds to it or contradicts it. What it doesn’t give you automatically is the thread itself.

The market that emerges when AI research tools reach mass adoption will be better at reading thermometers than any market in history. Faster, more distributed, more consistently calibrated to the available public data. And it will be collectively blind in a more systematic way: every thermometer reading the same temperature, none of them feeling the cold.

The farmer still has to decide it’s going to be a cold winter. And in a market where everyone’s instruments are measuring the same things at the same time, the structural risk isn’t that the instruments are wrong. It’s that when a new variable finally moves that was never in the model, the entire retail layer is wrong about it simultaneously.

A note on what we don’t know

We want to close with the same honesty we started with.

We don’t know whether the physical traders who actually set prices have already priced in the insurance mechanism and the restoration asymmetry. They have proprietary AIS data on every vessel. Analysts who follow P&I insurance markets specifically. They may have already arbitraged exactly what we’re describing, and the prices visible to everyone already reflect it in ways we can’t read.

What we can say is that the public forecast landscape has not adequately reflected these mechanisms. Whether that represents genuine mispricing or a gap between the visible and invisible layers of price discovery, we can’t tell from the outside.

What we’ve been doing, two people with a signal monitor and accumulated context, is closer to the farmer: spending enough weeks in the same field to notice when the pattern is breaking. Building the instruments to track it systematically. Being honest about what the instruments don’t measure.

The farmer buys the extra propane anyway. Because the cost of being wrong in one direction is much lower than the cost of being wrong in the other. And because the birds left early this year.