AI forecasting, explained: how AI predicts real-world events in 2026

AI forecasting means using AI to put numbers on the future: how much, how many, or how likely. In 2026 it works remarkably well in some areas, such as the weather, and is close to the best humans in others, such as real-world events. It cannot see the future. What it can do is turn the evidence that exists today into a probability that can be checked when the answer arrives. This guide covers how it works, how accurate it is, how it is judged and where it still fails.

Two kinds of AI forecasting

The term covers two quite different jobs:

• Forecasting quantities over time. How much electricity will the grid need tomorrow? How many units will a store sell next week? What will the temperature be in Oslo on Friday? Here AI learns patterns from long histories of data. This is the forecasting most businesses mean by "AI forecasting".
• Forecasting events. Will the Fed cut rates in December? Who will win the Senate? Will a ceasefire hold through March? These are one-off questions with a yes-or-no answer, and the forecast is a probability. This is the kind prediction markets trade and superforecasters are known for.

Both are judged the same way in the end: by comparing forecasts with what actually happened.

The clearest proof: the weather

Weather is where AI forecasting has made its most convincing progress. In 2023, Google DeepMind's GraphCast was more accurate than the leading conventional medium-range forecast on 90% of 1,380 test targets (Lam et al., Science). In 2024, its probabilistic successor, GenCast, was significantly more skilful than the world's top operational ensemble forecast on 97.2% of 1,320 targets (Price et al., Nature).

Weather is also where the most common scoring rule for probability forecasts comes from: the Brier score, designed in 1950 to grade weather forecasters.

How AI forecasts real-world events

Events are harder than weather: there is no physics to simulate, and each question is new. Most AI forecasting systems follow the same broad loop, much like a careful human analyst:

1. Take a clear question that will have a checkable answer by a fixed date.
2. Start from the base rate: how often have similar things happened before?
3. Research the evidence that exists today: news, data, official statements.
4. Check the sources and drop what cannot be verified.
5. Combine several independent forecasts, because independent errors tend to cancel.
6. Check the result against a track record, so that 70% forecasts come true about 70% of the time (calibration).

A system built this way is often called an AI superforecaster.

How accurate is AI forecasting in 2026?

On real-world events, close to the best humans, and improving fast:

• 2024: a research system that searched the news and combined several forecasts came close to the human crowd on competitive forecasting platforms (Halawi et al.). People who could consult an AI assistant forecast 24–28% more accurately than a control group (Schoenegger et al.).
• July 2026: on ForecastBench, several AI systems were "statistically indistinguishable from superforecaster-level accuracy," with caveats about the age of the human data (Forecasting Research Institute).
• September 2026: in Metaculus's spring tournament, professional forecasters still beat the top bots, but by just 1.25 points per question, a gap too small to be statistically significant (Metaculus).

We compare the benchmarks in AI forecasting benchmarks explained and the human-versus-machine debate in AI vs. superforecasters.

How AI forecasts are judged

A single forecast of 70% can never be right or wrong on its own. AI forecasters are judged across many questions:

• Accuracy with a proper scoring rule, usually the Brier score: lower is better, and 0.25 is what always saying 50% scores.
• Skill against a baseline, such as the base rate or the market's price. You can try both on your own forecasts with our free Brier score calculator.
• Calibration: do the 70% forecasts happen about 70% of the time?
• Timing: the forecasts must be made before the outcome is known. Tests on past events can be fooled by a model that simply remembers the answer.

Where AI forecasting still fails

• Remembering instead of forecasting. A model tested on events from before its training cutoff may already know what happened. This lookahead bias makes many AI backtests look far better than they are.
• The final hours. Near resolution, when news arrives fast, prediction markets still react faster than AI systems (Yang et al.).
• Brand-new situations. With no history to compare against, there is no base rate to start from.
• Overconfidence. Language models sound sure; without calibration, their 90% can behave like 70%.
• Prices that adapt. In the stock market, an edge disappears once it is widely used. See AI stock prediction: can AI predict the stock market?

A live example

The only test that really counts is forecasting events before they happen, in public, and then being scored. Sikt Intelligence, which builds an AI superforecaster, is doing that on the 2026 US midterms:

A live AI forecast, scored in public. In the Sikt Midterm Bench, Sikt forecasts the 2026 midterms next to Kalshi and Polymarket. Latest round, Oct 3, 2026: Senate: Democrats 56% (markets 64% to 66%), every round; House: Democrats 88% (markets 91% to 92%), every round. Every forecast will be graded after election day, misses included.

Key takeaways

• AI forecasting covers both quantities over time, like weather or demand, and one-off events, like elections.
• In weather, AI now beats the leading conventional forecasts on most targets.
• On real-world events, the best AI forecasters are close to top humans in 2026: tied on one benchmark, narrowly behind on another.
• Judge any AI forecaster by proper scores and calibration on questions that were in the future when it forecast them.
• Its main weaknesses are lookahead bias in tests, speed near resolution, novel situations and overconfidence.

FAQ

What is AI forecasting?

Using AI to estimate future quantities, such as demand or temperature, or the probability of future events, such as an election result. Event forecasts are probabilities that are checked against what actually happens.

How accurate is AI forecasting?

In weather, AI models now beat the leading conventional forecasts on most targets. On real-world events in 2026, the best AI systems are statistically tied with superforecasters on ForecastBench and narrowly behind professional forecasters on Metaculus.

Can AI predict the future?

Not with certainty, but it can estimate the odds of real events close to the level of the best human forecasters. We cover the evidence in can AI predict the future?

How is an AI forecast evaluated?

With a proper scoring rule such as the Brier score across many questions, a calibration check, and a comparison with a baseline such as the market's price, using only questions that were in the future when the forecasts were made.

Sources

• Lam et al. (2023): GraphCast: Learning skillful medium-range global weather forecasting, Science
• Price et al. (2024): Probabilistic weather forecasting with machine learning, Nature
• Halawi et al. (2024): Approaching Human-Level Forecasting with Language Models
• Schoenegger et al.: AI-Augmented Predictions: LLM Assistants Improve Human Forecasting Accuracy
• Forecasting Research Institute: AI models have likely reached parity with superforecasters (July 2026)
• Metaculus: FutureEval spring 2026 tournament results (September 2026)
• Yang et al.: LLM-as-a-Prophet: Understanding Predictive Intelligence with Prophet Arena