Not subscribed? Sign up to get it in your inbox every week.

A bot won a forecasting tournament in September.

The Summer 2026 Metaculus Cup had 885 forecasters working through 58 questions, on things ranging from the price of oil to limits on data-center construction. The #1 and #2 performers were bots (Laertes and ManticAI), with the best human forecasters placing #3 and #4, and FutureSearch's bot took 5th. There are plenty of caveats to any single season, and I'll get to them, but I think this points at the biggest shift coming to data teams.

So much of data analysis is written in the past tense: "What was churn last quarter?" or "What drove pipeline?" or "Why did margin decline in March?" AI analysts sped up the process for answering those questions, letting you type the question in English and get the query back in seconds, but the answer still describes something that already happened. The questions business decision-makers are actually facing are written in the future tense. Will this customer renew? Will that competitor cut prices? Will the new tariff hit our supplier before we can switch?

I predict that in just a few years, every business will have its own superforecasting agents, able to use its internal data alongside deep research to answer those questions. The building blocks for that agent are already on the market.

Superforecasting is a term that comes from Philip Tetlock, describing the people in his Good Judgment Project who reportedly performed about 30% better than intelligence analysts with access to secret information. Much of their edge came from habits you can write down:

A superforecasting AI is an agent that runs this process. Cassi, one of the top systems on the main public benchmark, breaks the question down into sub-questions, generates search queries for them, reads sources and determines which are relevant and up to date, then makes a prediction. Mantic tests its system on historical data, then grades and improves it. In the end, you get a numerical prediction along with the reasoning and sources.

This is different from the kind of forecasting your finance team might do, where you fit a time series to previous data points and extrapolate. That's good for things like next month's demand, as long as next month resembles the past. Superforecasting AIs are designed for one-off questions with no clean history, which they handle by reading everything around them.

In January, the Forecasting Research Institute (FRI) extrapolated that AI would match human superforecasters around November 2026. In July it reported that multiple models had likely matched superforecasters on its ForecastBench benchmark. Similarly, Metaculus's team of bots trailed its pro forecasters by 20 points in a head-to-head competition in Q2 2025, but only 1.25 in spring 2026.

On that last measure, humans still win. Also keep in mind that FRI's comparison leans heavily on extrapolation, since its human panel was last surveyed in 2024, and that the Cup's scoring reportedly favors forecasting early and often, which plays to the bots' strengths. Saying AI forecasters are now roughly even with the best humans seems fair.

The cost is a different story. According to The Economist, a report from human superforecasters can cost over $10,000 and take a week, while FutureSearch quotes a few dollars and a few minutes. Where an organization might have commissioned a few reports a year on key decisions, it could now run one on every deal in the pipeline.

Everything in the above examples comes from publicly available information. Sure, you could ask a system like this what the price of oil might be, but it knows nothing about your renewal with Acme. FutureSearch's own docs mention that you need to pass in things like company size, budget, and what you've tried before, because the forecaster can't look them up.

Lots of the data that would help answer those kinds of questions sits inside companies (product usage, CRM history, support tickets, supplier emails), and it's the same data you're already writing queries against. Therefore it can feed a different set of questions:

The most interesting part here is that forecasts come with their own answer key: when December 31 arrives, you know whether Acme closed or not. I argued in #153 that review tasks are the first place to tune your own model because the task itself labels the data, and forecasting is similar. Lightning Rod puts this to work, turning historical data into resolved questions without hand-labeling. It claims, on its own evaluation, a 34% lower Brier score (more on that below) than GPT-5 at predicting supply-chain disruptions.

Over a few years, that adds up to a scored record of what your company saw coming and what it missed. Anyone can buy the public forecast, but you will own the history, which is the same argument I made in #147 about owning your intelligence stack.

It also changes the data team's job. The value shifts away from doing the research and modeling itself and towards framing questions that can be answered cleanly, managing the source-of-truth data, ensuring its trustworthiness, and analyzing misses.

I can't find anyone who sells all of this today, so what follows splits into two halves. Consider them all specimens of what's coming rather than recommendations, since most of the accuracy claims are self-reported.

Tools for researching the world:

Tools for working from your internal data:

The internal superforecaster is what you get when those two halves merge: an agent that reads the world and your warehouse, commits to a number, and gets graded on it.

A few problems are worth knowing before you get excited.

Small sample sizes. A public benchmark can score thousands of questions, but your company may only resolve a few dozen a year. It may take a while for your agent's score to be meaningful, so don't read much into one or two questions.

Actions affecting the outcome. Your agent might predict that Acme has a 20% chance of renewing, so you have your customer success team reach out, and Acme renews. Did your agent make an accurate prediction? To keep track of this, log any actions you take alongside each forecast.

Confidential information shared with third parties. To get the most out of the tools that research the outside world, you'll likely need to provide some private context, and that context goes to a third party. Start with questions where you're comfortable sharing the information, and hold these agents to the controls laid out in #154.

Human decision-makers. Your agent will give you probabilities, but it's up to you to decide what action (if any) to take. That means weighing both the probability of an event and how damaging it would be: a 30% chance of losing your top supplier could mean doing nothing or starting to dual-source today. Always make sure a named human is responsible for the call, and that they never hide behind "the agent said there was a 70% chance."

What are five arguments happening on your leadership team right now that will be resolved in the next 90 days? Write them down as yes/no questions with dates and a specific source of truth for each.

Have the person responsible for each question write down a probability privately, then run the questions through FutureSearch (or a frontier model with web research) with a prompt like this:

Put both numbers in a spreadsheet, then go back and score each question when it resolves. The Brier score works well for this: take (probability - outcome)², using a decimal for the probability (0.8 instead of 80%) and 1 for something that happened or 0 for something that didn't. Lower is better, and someone who said 50% for everything will get a score of 0.25.

In 90 days you will have a number for your agent and a number for your leadership team. You've very likely never seen the second one.

Till next time,
Chris