The Art of the Prediction

Nobody can predict what Trump will do next. That’s the consensus, and it’s not unreasonable. He withdrew from Afghanistan, then deployed to Venezuela. He threatened to buy Greenland, then dropped it over lunch. He imposed the most sweeping tariffs in a century, then reversed them in a week. His own advisors have said, on background, that they often don’t know what’s coming.

As someone who has spent thirty years working in crisis management and public affairs, I’ve long believed our industry should be able to do better than this. Governments set policy, investors allocate capital, and businesses plan strategy based on assumptions about what the US president will do next. When those assumptions are wrong — as they were for most analysts on Liberation Day — the cost is measured in billions, broken alliances, and months of misallocated effort.

For most analysts, this unpredictability is where the analysis stops. It’s treated as a feature of Trump’s personality — chaotic, impulsive, impossible to model. The implication is that anyone trying to anticipate his moves is wasting their time.

We think this is wrong. Not because Trump is predictable in any conventional sense — he isn’t — but because the unpredictability may mask patterns that only become visible when you look at every crisis simultaneously, hold them all in view at once, and ask what’s genuinely consistent across every single one.

No human analyst can do this. The cognitive load is too great. You anchor to the most dramatic case, you overweight the most recent one, you unconsciously downweight the ones that don’t fit your thesis. These aren’t failures of intelligence. They’re heuristics — features of how human brains work under complexity.

But what if you combined human intelligence — the contextual knowledge, the professional judgment, the ability to explain why something works, not just that it works — with machine intelligence that can hold a dozen complex episodes in perfect parallel, apply identical questions without drift, and report findings without hedging?

We tried it. We built a model. And it works.

Across every crisis we studied, the same small set of drivers kept appearing — and a much larger set of factors that most analysts treat as important turned out not to matter at all. What moves this president is not what conventional analysis assumes. The signals that actually predict his decisions are more personal, more immediate, and more consistent than the strategic frameworks that dominate the commentary.

Get the drivers right and the trajectory becomes readable. Miss them, and you’re watching the wrong screen.

What the Model Got Right

We have tested the model against every major crisis of Trump’s second term. In every case, it identified the trajectory before events confirmed it — and told us which signals to watch and which to ignore.

Liberation Day tariffs (April 2025). The stock market crashed 12% in four days. Trillions erased. Trump posted “BE COOL!” on Truth Social. Most analysts expected the equity crash to force a reversal. It didn’t — and our model explained why. Equity markets are abstract. They don’t generate the kind of personal, visible pressure that moves this president. What the model identified as the real trigger was the bond market: when Treasury yields spiked and the 30-year market began to dislocate, the pressure shifted from abstract market pain to something that threatened the functioning of the US government’s own borrowing. When that trigger fired on April 9, the 90-day pause followed within hours. The S&P surged 9.5% — its largest single-day gain since 2008.

The model got the direction right and identified the correct mechanism. What it didn’t get was timing — a recurring limitation we’re working to address.

Venezuela (January 2026). The model predicted no retreat. It was right — though the operation concluded faster than we expected. Congressional concern, international condemnation, Russian and Chinese protests were all present. The model identified in advance which of these signals would influence the president and which wouldn’t.

Greenland (January 2026). The model predicted a maximalist opening followed by retreat to a face-saving “framework.” That’s what happened at Davos. Eight nations issued a joint statement. Nothing changed. One person sat across a table and offered Trump something he could call a deal.

Minneapolis (January 2026). Two civilians killed by federal agents during immigration enforcement. The first killing did not trigger de-escalation. The second one did. On the surface, these events look similar. The model’s framework helps explain why they produced different responses — though we’d be overstating our confidence to say it predicted the precise sequence in advance.

Iran: What the Model Said

We scored Iran three times in one week. Each time the inputs changed. Each time the output moved.

After the Geneva talks on Tuesday, we scored it at roughly 60% probability of a deal. After Trump’s deadline on Thursday, the probability held but the window narrowed. Then the Supreme Court struck down his tariffs on Friday morning — a 6-3 ruling, personal and humiliating, delivered by justices he appointed — and everything shifted.

Our prediction at the time: approximately 70% that a framework deal would emerge. The tariff defeat didn’t just wound the president politically. It created urgent pressure for a compensating victory in another domain. Iran — where a framework was already forming and the trophies were already on the table — was the obvious stage.

How this prediction resolved: see Edition 2.

Cuba: What’s Coming

The model scores Cuba among the highest of any scenario we’ve tested. Every driver is present and strong. The grievance runs deep — Obama’s Cuba normalisation was a signature policy Trump reversed in his first term. “Freeing Cuba” would be the most dramatic US foreign policy headline since the fall of the Berlin Wall. The transactional value includes cobalt and nickel reserves, real estate ninety miles from Florida, and permanent denial of Russian and Chinese influence in the Western Hemisphere. With Rubio as Secretary of State, the usual institutional friction isn’t just absent — it’s actively pushing toward escalation.

The pressure campaign is already underway. Venezuelan oil shipments have been cut. Trump signed a national emergency declaration on January 29 imposing tariffs on any country supplying oil to Cuba. Air Canada has cancelled flights. Trump told reporters Cuba is “a failed nation” that should “absolutely make a deal.” Rubio has stated his only topic of discussion with Havana is when the regime will relinquish power.

The question is not whether Cuba escalates. The question is the form and the timing.

The Art (and the Science) of the Prediction — Edition 2

Last week we published our second Substack article. In it, we predicted a 70% probability that a framework deal would emerge between the United States and Iran. We explained which signals to watch. We described what would confirm or contradict the trajectory.

Two days later, the United States and Israel launched Operation Epic Fury.

Khamenei is dead. Six American service members have been killed. The Strait of Hormuz is effectively closed. Oil is past $79. And the world is watching a military operation on a scale not seen in the Middle East since 2003.

We were not alone in expecting a deal. Oman’s foreign minister declared a “breakthrough” on the morning of February 27 — the day before the first strikes. Polymarket bettors gave the US strike only a 23% probability by the end of February, as recently as February 9. Iran’s foreign minister was telling reporters that the Geneva round was the “longest and most serious” yet. The entire diplomatic apparatus was pointing toward agreement.

Almost nobody called this. Certainly not at this time, and not at this scale.

So we got it wrong. The question is whether we got it wrong in a way that tells us something useful — or whether the model simply failed.

What happened

Let us start with the facts as we now understand them.

The operation had been planned for months. Israeli military officials have since confirmed that the US and Israel built an extensive target bank while diplomatic talks were continuing in parallel. The Washington Post reported weeks of lobbying from Saudi Arabia and Israel to move Trump toward military action. Secretary of State Rubio’s account on Monday was revealing: the US knew Israel was going to strike, knew that would trigger retaliation against American forces, and concluded it was better to act jointly.

Congress was not consulted. The loyalist cabinet offered no pushback. Defence Secretary Hegseth told reporters that Trump has “all the latitude in the world” to determine the timeline.

In other words: the diplomatic track and the military track were running simultaneously. We — and most observers — read the diplomacy as the dominant signal. The more accurate reading is that both tracks were genuine: had diplomacy produced what the US was demanding, the military option would likely not have been actioned. But given the scale of those demands, the administration was always prepared for diplomacy to fail — and was building the military option in parallel.

What the model got right

The model identified Iran as one of the highest-scoring scenarios we have tested. Every driver was present at near-maximum levels. The personal motivation, the headline value, and the prize on the table were all enormous — and that was true whether Trump pursued a deal or a strike.

The model got the question right. It got the drivers right. It identified the correct signals to watch. It told us that something very large was about to happen on Iran.

What it got wrong was the form.

What the model got wrong

Our model includes a factor that captures how much effective friction exists within the system — the institutional pushback, the political costs, the practical obstacles to any given action. We assessed that friction against a strike was significantly higher than friction against a deal: Congressional dynamics, allied reluctance, Treasury concerns about oil disruption, the institutional memory of Middle Eastern wars.

That assessment was wrong. The friction against a strike was dramatically lower than we scored it, because the factors that lowered it were invisible from open sources. Months of military planning. Saudi-Israeli lobbying conducted behind closed doors. A cabinet already aligned. Congress bypassed entirely.

When we correct that single input — keeping everything else identical — the strike overtakes the deal as the highest-scoring pathway. The framework was sound. The input was wrong.

That is an honest assessment, not a consolation. The practical lesson is clear: when scoring how much resistance exists against a given action, we should present a range rather than a single number, and explicitly identify what might be happening beneath the surface. We are building that into the model now.

Same model, corrected input: the strike overtakes the deal.

Venezuela changed the president

This may be the most important finding from the Iran case.

Trump’s first-term record showed consistent reluctance toward large-scale military force. He pulled back from striking Iran in 2019 over estimated casualties. He withdrew from Syria. The Soleimani strike was targeted and self-limiting.

The second-term record is different. Venezuela — a 150-aircraft leadership change operation that produced zero US fatalities, Congressional war powers resolutions that failed, and a Medal of Honor ceremony at the State of the Union — demonstrated to this president that maximalist military action works, that the costs are manageable, and that the political system absorbs it.

The man who ordered the Iran strikes is not the man who called off a planned strike on Iran in 2019, ten minutes before execution, over estimated casualties. He has been changed by what happened in Caracas. Our model needs to account for that shift, and we are working on it.

The Venezuela comparison has limits, however. Iran is a fundamentally different adversary: a country of 88 million with a battle-hardened military, deep ideological commitment at the leadership level, and a proven willingness to absorb punishment and fight on. Venezuela’s leadership folded; Iran’s will not, at least not quickly. The probable common end state — a leadership smaller and sufficiently cowed to concede on nuclear enrichment, ballistic missiles, and the axis of resistance — may look similar in outline, but the path to get there is far longer and far costlier.

The 72-hour signal

Here is what has been most striking since Saturday.

Within 36 hours of the first strikes — while bombs were still falling and American service members were dying — Trump told Axios he had several “off-ramps.” He told The Atlantic he had “agreed to talk” to Iran’s new leadership. He framed the strikes explicitly as punishment for refusing the deal: “They should have done it sooner. They played too cute.”

That is not the language of regime change. It is the language of someone already building the exit. And the military reality reinforces this: Iran is simply too large to invade, and the force ratios in the Gulf are nowhere near sufficient for a ground campaign. Air and missile strikes can destroy infrastructure and kill leaders, but they cannot occupy a country of 1.6 million square kilometres. The exit is not just politically desirable — it is operationally inevitable.

Deal vocabulary appeared within 36 hours — while strikes were still intensifying.

We have been tracking every public statement Trump has made since the strikes began. The pattern is unmistakable: dominance language dominates Saturday. By Sunday afternoon, deal language appears. By Monday, the two are running in parallel.

Our model identifies specific mechanisms that govern how this president’s actions evolve once initiated. One of them is the construction of a narrative that allows a change of course to be presented as victory. That mechanism is already active. Trump can now say: I gave them every chance. They refused. So I destroyed their military, killed their leader, and now the new leadership is ready to talk. This is the deal Obama could never have gotten — because I showed strength first.

That narrative — a deal reached through demonstrated overwhelming force — is more politically valuable than a quiet diplomatic agreement ever would have been.

Our assessment, shared by our intelligence partners at Harquebus: the deal did not die on Saturday. It changed shape. The trajectory will ultimately bend back toward negotiation — but from a position of devastating military advantage rather than diplomatic compromise.

The costs are building

Every observable input trending toward higher friction.

Six US dead. Oil past $79. Markets falling. Hormuz closed. Congressional war powers votes this week. GOP Representatives Massie and Davidson publicly opposing. UK cautious. EU split. Iran’s interim leadership saying they “will not negotiate” — though the model suggests that position will shift as the asymmetry of costs becomes clearer.

Every day the operation continues, the friction against it increases. Our model predicts that when cumulative costs cross a threshold — when continued strikes become more politically expensive than a deal — the exit pressure reasserts.

The clearest leading indicator is language. Track the ratio of dominance vocabulary to deal vocabulary. When deal language consistently outweighs escalation language, the trajectory is shifting. It has already begun.

What would change this assessment

Two things.

First, if the conflict takes on its own momentum. A mass-casualty event against a US carrier or base. A sustained energy crisis from the Hormuz closure. Direct ground engagement. At that point the assumption that the decision-maker controls the exit timing no longer holds.

Second, if Trump commits to regime change as the explicit, sustained objective — no deal with any Iranian government — and absorbs costs well beyond anything we have seen him tolerate. That would be genuinely new territory.

Neither has happened yet. The signals from the first 72 hours — “off-ramps,” “agreed to talk,” “limited in scope,” “four to five weeks” — all point toward a president constructing his exit while the operation is at peak intensity.

What comes next: Cuba and beyond

Cuba remains the highest-scoring scenario we are tracking. Every driver is present at near-maximum. The economic squeeze is already underway — Venezuelan oil shipments cut off, national emergency tariffs on alternative suppliers, flights cancelled, the economy grinding to a halt. Rubio’s personal investment in Cuba policy means the usual institutional friction is not just absent but actively pushing toward escalation. Once Iran resolves — and our assessment is that it will resolve in weeks, not months — Cuba moves to the front of the queue.

Panama has largely played out as the model predicted. Resistance collapsed early. The trophy is being delivered without confrontation.

NATO remains a slow burn. The model scores formal withdrawal as unlikely but scores the hollowing-out of Article 5 through bilateral deals and conditional security language much higher. Watch the 3–5% GDP defence spending demands.

What we are building

The Iran case exposed a genuine limitation: our model scored a single point estimate for how much resistance existed when a range would have been more honest. We are building structured protocols for this — presenting visible inputs alongside an explicit assessment of what might be invisible.

We are also developing a real-time monitoring tool that tracks the observable signals daily: casualty counts, oil prices, Congressional statements, Trump’s language patterns, adversary signalling. The aim is to move from periodic predictions to continuous tracking with clear thresholds. We will share more on this in the coming weeks.