The AI That Saw the Iran Strikes Coming · Anthony Vinci, Vico

What we talked about.

Anthony Vinci of Vico returns to the show to explain how his company's forecasting engine, the Vico model, called the Iran strikes well before markets did. While prediction markets sat near 30 percent in mid February for strikes by the end of March, the model climbed from 60 and 70 percent to 87 percent days before the event. It also flagged the closure of the Strait of Hormuz about 10 days out. Vinci describes how the system steadily weighs facts like carrier movements and official statements while discounting rhetoric, and he admits misses, including a short-lived peace agreement. He explains why he now trusts probabilities over his gut, citing a North Korean missile test the model predicted correctly. The conversation covers the US and China summit forecast, where prediction markets fall short, current watch items such as another possible Iran strike and Federal Reserve rate moves, and new tools called Scenario Explorer and Causal Explorer. Vinci also predicts that within a decade, big decisions will routinely come with a probability attached.

An AI forecast US strikes on Iran at 87% while prediction markets sat near 30%. Anthony Vinci explains how it saw it coming.

The questions, and the answers.

What did the AI see on Iran that the markets missed?

Markets were fence sitting, caught up in rhetoric and arguments for and against. Our model systematically added up each new fact, like an aircraft carrier moving or a presidential statement, weighed it against everything else, and stepped its forecast up over time. It doesn't care about the discussion. It just adds up information, and that's how it called it.

What does the model get wrong, and how do you handle that?

We missed a so-called peace agreement with Iran, and you can see that in our newsletter. The model didn't take it seriously, which played out to be true, since within days or weeks they were striking each other again. Prediction markets focus on the words because people have bets. We focus on practical reality, like whether oil will actually flow.

Has the model ever changed your mind on something you felt sure about?

All the time. Early on it said there was an 80 or 90 percent chance of a North Korean missile test within 90 days, and I didn't believe it. It turned out they test every 90 to 120 days and we were about 85 days into the cycle. It was right. Now I check the model first, though I still form my own opinion.

If you acted on every call it made, would you be ahead or broke?

We'd be ahead. In our forward testing, which covers things we actually published and not backtesting, we're getting 85 to 90 percent of calls correct, and in some cases higher than 90. We're ahead of prediction markets too, so there's clearly alpha in that, and we're looking into what it means for trading.

How does the model decide when a leader says one thing and does another?

It writes out its reasoning. It might note that a person said this on a given date, and that they've said similar things seven times in the past nine months without following through, so it discounts the statement. It can assign something like a 2.5 percent increase in probability. Humans can't remember every statement a politician made over years, but the model can.

Where do prediction markets beat you, and where do you beat them?

They're strong in high volume areas like presidential elections and economic questions such as inflation, where wisdom of the crowds works. They're weak in low volume markets, like a small town mayoral race or most international elections. They also struggle with wars and long buildups, where people rely on a couple of news sources. The model reads every language and ignores rhetoric.

What problem does your decision support tooling solve?

Organizations face so many factors that tracking them is hard. A typical company plans for maybe three to five scenarios. With Scenario Explorer you can play out hundreds or thousands, and we did one with 3,500. It's like the Marvel multiverse. Causal Explorer then shows the causation behind the odds, so you can test what happens if something changes.

In 10 years, will anyone still make a big call without a probability next to it?

I don't think so. We'll look back at getting ten old men in a room to argue while the most senior person picks one, and see it as one step away from witch doctors. All organizations will use some form of decision support tool, and I think individuals will too, even for things like choosing a college.