On 3 December 2025, The Forecasting Machine was asked whether Nicolás Maduro would be removed as Venezuela’s president before 31 January 2026. It answered with a probability of about one in four. On 3 January 2026, US forces captured Maduro in Caracas and flew him to New York to face charges. On 17 February, Arthur Balick, one of The Forecasting Machine’s co-founders, wrote up the forecast on LinkedIn.
This piece looks at what the machine was weighing at the time and what happened. It also explains why one correct-looking call is weaker evidence than it seems.
The 2025 machine forecast also had a public crowd-market predecessor. On 30 July 2024, Servan-Schreiber shared a Hypermind screen that put only 17% on Maduro losing power before 2025 and 83% on his remaining. Maduro stayed through that deadline, so the crowd’s favoured outcome occurred. The later question covered a different window and a much more acute military situation; the two numbers should not be read as forecasts made from the same evidence.
The situation on 3 December
By early December the pressure on Caracas was visible. According to the House of Commons Library, the United States had built up forces in the Caribbean from August 2025. It had raised its reward for information leading to Maduro’s arrest to $50 million and, from September, struck boats it said were carrying drugs. The USS Gerald R. Ford carrier strike group entered the Caribbean on 16 November. Stars and Stripes described it as the largest US military presence in the region in decades.
Al Jazeera reported that Donald Trump and Maduro spoke by phone on 21 November. Citing Reuters sources, it said an offer of safe passage out of the country had come with a deadline that expired on 28 November. On 29 November Trump declared Venezuelan airspace closed. Maduro had gone several days without a public appearance before he reappeared at a rally on 1 December and rejected what he called a “slave’s peace”.
How the machine broke the question down
Balick’s post describes three routes by which Maduro could leave power within the window:
- a US military intervention;
- a palace coup by insiders, triggered by the threat of invasion;
- a negotiated exit.
Splitting a question into separate pathways, then combining their probabilities while allowing for the fact that they are linked, is a standard forecasting technique. Balick lists it among the method’s ingredients, together with “reference class forecasting”, which means starting from how often similar situations have ended in a given way.
Balick wrote that the machine had picked out a few high-information signals from a stream of mostly noisy Venezuela news. He cited a Caribbean deployment he said exceeded the 2011 Libya Tomahawk build-up, the Ford’s positioning, Maduro’s four-day disappearance, rotations in his security detail and fractured loyalty in the officer corps. These are his descriptions of the machine’s inputs; we have confirmed the deployment, the carrier’s arrival and Maduro’s absence from public view, but not the claims about his security or the officer corps.
He also contrasted the 25% with what he described as a market-implied probability below 12.5%. He did not say which market. His point, aimed at insurers, is that pricing a risk at half its real likelihood is a costly error.
What happened
The pressure kept building through December, with the seizure of a tanker and a blockade on sanctioned oil tankers. On 3 January 2026, in what the US Department of War called Operation Absolute Resolve, US forces raided Caracas with more than 150 aircraft and captured Maduro and his wife, Cilia Flores. The department reported no US deaths. Cuba said 32 Cubans were killed. The couple were taken to the USS Iwo Jima and then to New York, where they pleaded not guilty on 5 January to charges including narco-terrorism conspiracy. Delcy Rodríguez was sworn in as interim president the same day.
The route turned out to be the first of the three, military intervention, and it came four weeks before the deadline.
Why one hit is not a track record
A 25% forecast means the event should happen about one time in four. It does not say whether this case will be the one, so an event forecast at 25% that does happen is neither a vindication nor an error. Had Maduro stayed, the forecast would have been just as defensible.
Forecasts are judged in bulk. The test is calibration: across many questions given 25%, roughly a quarter should come true, and the same goes for every other level. Accuracy scores such as the Brier score, the average squared gap between the probabilities given and what actually happened, reward forecasters who are both calibrated and willing to move away from 50% when the evidence allows.
