Mantic is like a deep research agent that's built specifically for forecasting. It uses LLMs throughout the pipeline, first doing research on the question and then reasoning towards a prediction. Alongside LLMs, the system also uses more traditional methods for modelling data.

Currently, the system is powered by a mix of Gemini, Claude, and OpenAl's reasoning models. We are working on post-training open source models to make accurate predictions.

We are constantly iterating on the system to improve its accuracy, as tested on datasets of real-world forecasting questions.

Mantic currently has access to:

• Licensed and open news data
• Wikipedia
• Economics: a broad range of quantitative datasets
• Finance: market price data across many equities, forex
• SEC filings
• Predictions from other platforms like Metaculus and Polymarket
• User-provided links

We're working to expand this list.
Mantic does not currently have access to the open web.

We're constantly improving Mantic's accuracy. This is the main focus for our R&D.

As far as we know, Mantic is currently the most accurate general-purpose Al forecasting system in the world. We won the highest prize money in the Metaculus Al Benchmark Tournament for Q1 2025, and have made continued progress since then. The Metaculus competition is the premier benchmark for Al forecasting agents. They release new questions daily, totalling several hundred high-quality questions per quarter. We don't think our current system could outperform a top human forecaster, but it would likely outperform an amateur.

Topics To Avoid

• Private life:
  Mantic won’t know anything about your private life, so can’t give good predictions here.
• Stock prices:
  Predicting stock prices is too difficult.
• Long-term:
  Mantic has been developed on predictions up to only about 6 months into the future. You can ask longer-term questions but we don’t know if the predictions will be strong.

Good

• Work-relevant:
  We’d love users to ask questions that are relevant to their work.
• Example topics:
  New technologies, policy changes, competitor behaviour.
  Think: what could I read in the news that would affect my work?

Wording

• Probability:
  You don’t need to say “What’s the probability that…” Just ask what you want to know, e.g. “Will…” The answer is always probabilistic.
• Deadlines:
  Prefer rolling time windows over fixed calendar deadlines.
  For example, instead of “Will the President do X before 1st December?”, it’s better to ask:
  “Will the President do X in the next 6 months?”
• Outcomes:
  You can’t ask open-ended questions (e.g. “Who will be the next British MP to resign?”).
  For a multiple choice question, you’ll need to provide the set of possible answers.
• Specificity:
  The questions should be as specific as possible.
  For example, instead of asking whether any country will default on its sovereign debt, you could ask for a specific country like Argentina.
• Hypotheticals:
  You can technically ask conditional questions (e.g. “If X happens, will Y happen?”) but we haven’t yet focussed on this so it might misfire.

Supplementary Information

• You can include information you think Mantic won’t have access to (see here for what data sources it has).
  This can be text, or you can paste a URL.

Clarifications

• After the fact, it should be easy to judge what the answer to the forecasting question was.
  Therefore, you can add clarifications where you rule in or rule out certain scenarios, e.g.
  “If the President is impeached, that’s not alone sufficient to count as ‘No’.”

Public or Private

• All questions are anonymous, even public ones.
  Public just means that other users can find the question through search.
  We encourage public questions, so that other users can benefit.

Multiple Choice Questions

• MECE:
  You can define the multiple choice outcomes, but these must be mutually exclusive and collectively exhaustive.
  In other words, they shouldn’t overlap and they should cover the whole space of possibilities.
  You might need to add an “Other” outcome.

Numerical Questions

• Definition:
  Numerical questions are where the answer is a number, e.g. “How many cars will Tesla produce in Q3 2025?”
  Mantic gives a probability distribution over the different possible numbers.
• Range:
  You specify the range of possible values, e.g. 300,000 to 400,000.
  It’s OK if the answer might fall outside of this range, because you can specify that the lower bound and/or upper bound is open-ended.
  If so, Mantic’s forecast will accordingly include the probability that the number is above or below the range.

‍