A Competition in Memory of Enrico Fermi, Promoting Scientific Thinking and Confidence Calibration

The Fermi Competition trains three cognitive skills rarely taught: Fermi estimation (breaking complex problems into smaller parts), calibration (accurately assessing your own uncertainty), and Bayesian updating (revising beliefs in the light of new information). Students answer estimation questions, assign confidence levels to each answer, and at halftime receive hints that let them update their estimates.

These are skills used by intelligence analysts, quantitative traders, and professional forecasters. The Fermi Competition makes them accessible to students at the secondary and university level.

Yes. There is no cost to participate because we are generously supported by philanthropic donors. We are always on the search for new sponsors and donors to help us grow.

The official competition launches in Spring 2026. The provisional date is April 16. Practice mode and guest mode are available now, so you can start preparing immediately.

Practice mode can be run by teachers for their entire class and is highly recommended before the official competition. Teachers sign up and distribute credentials to students, just like in the official competition. It's a great way to familiarize students with the format, timing, and the unique challenge of combining Fermi estimation with calibrated confidence.

Guest mode can be started without signing up or creating an account. It's a quick way to try the competition, but no results are stored.

The competition is 50 minutes. At the 25-minute mark, hints are revealed for each question, giving you additional information that may help you refine your estimates and adjust your confidence levels. You can finish earlier if you wish. Schools should allow some buffer time for setup and instructions.

There are up to 20 questions. You don't have to answer all of them. You can attempt many questions quickly with rough estimates, or focus on fewer questions with careful reasoning. Neither strategy is obviously better. It depends on your strengths and how confident you are in your answers.

One useful approach: if you encounter a question where you're struggling to make good assumptions or find a way to break it down, consider moving on and returning to it after the halftime hints are revealed. The hints often provide exactly the kind of anchor or starting point that makes a difficult question tractable.

A Fermi question asks you to estimate something that seems unknowable but yields to structured reasoning. Example: "How many dentists work in the US?"

You don't guess randomly. You decompose the problem into smaller, estimable parts. US population is roughly 300 million. How often does the average person visit a dentist? Maybe 1.5 times per year. That gives about 450 million dental visits annually. How many patients can one dentist see? Perhaps 15 per day, working 200 days per year, so roughly 3,000 patients annually. Divide 450 million by 3,000 and you get about 150,000 dentists. The actual answer is around 200,000. The estimate is between half and double the true value, so it's deemed correct.

Several techniques help:

Decomposition. Break the problem into parts you can estimate separately, then combine them. Each piece should be something you can reason about.

Reference classes. Ask: what's a similar quantity I already know? If you're estimating the number of hospitals in a country, you might anchor on knowing roughly how many exist in your city or state, then scale up.

Rounding. Use round numbers. 300 million is easier to work with than 340 million, and the small errors often cancel each other out. Fermi estimation is about getting in the right ballpark, not precise arithmetic. Round aggressively and keep the math simple.

Bounding. Think about what would be implausibly high and implausibly low. There can't be 3 million dentists in the US (that would be 1 in 100 people). There can't be only 1,000 (that would mean each dentist serves 300,000 people). Your answer should fall somewhere between these bounds.

Sanity-checking. After calculating, ask: does this make sense? If your answer implies something absurd (every household owns 50 pianos, or one dentist serves an entire metropolis), you've made an error somewhere.

Yes. The competition tests numeracy, but numeracy means far more than fast arithmetic. You can be quick with mental calculation and still be somewhat innumerate if you lack intuition for orders of magnitude, can't distinguish meaningfully between millions and billions, or don't notice when an answer is off by a factor of 1,000.

The skills that matter here are higher-level: making good assumptions, decomposing problems, examining your mental model of the world, and calibrating your confidence. A calculator doesn't materially help with any of that. It just removes a distraction so you can concentrate on what actually matters.

And if you round aggressively, you won't need a calculator much anyway. Multiplying 100 million by 2 and dividing by 100 is straightforward mental arithmetic. With some basic proficiency in mental calculation, you might even be faster without a calculator, since you skip the overhead of typing numbers in. Having or not having a calculator is not a big deal either way.

An answer is correct if it falls between half and double the true value. If the actual answer is 200,000, anything from 100,000 to 400,000 is deemed correct.

You assign a confidence bucket to each answer:

Everyone starts with 250 base points, so you can't go negative even if everything goes wrong.

The 0–20% bucket is special: +3 if right, 0 if wrong. There's no downside for admitting you're genuinely uncertain. Knowing that you don't know is a skill, and we don't punish it.

Skipping gives 0 points. There is no penalty, but even a wild guess at 0–20% has positive expected value.

This is a proper scoring rule: the math guarantees that honest reporting maximizes your expected score. If you're uncertain which bucket fits, pick whichever feels closest to your true belief.

Calibration is the skill of accurately knowing what you know. A well-calibrated person who selects the 60–80% confidence bucket should be right somewhere between 60 and 80 times out of 100.

Even trained professionals struggle with this. In a study by Professor Jeffrey A. Friedman, national security officials from more than forty NATO allies and partners were overwhelmingly overconfident. When officials said there was a 90% chance that a statement was true, those statements were actually true just 57% of the time.

Calibration is trainable, and this competition trains it through repeated practice with clear feedback. After competing, you'll see your calibration curve: stated confidence vs. actual accuracy. If you selected 60–80% and were right only 25% of the time, you're overconfident at that level. This feedback is rare and valuable.

Why train low confidence (0–20%, 20–40%)? Most calibration tools start at 50% because they use binary yes/no questions. Fermi estimation is different. You can't flip your answer. You might be genuinely, irreducibly uncertain. Learning to express "I don't know" with appropriate confidence is an invaluable skill almost never taught.