Forecasting the median
How many people will be admitted to hospital with influenza in the United States in the week ending Saturday, Nov 14, 2026?
This project maintains a community forecast. As a prediction contest, participants earn points based on how much they improve the forecast.
It is a median, meaning the community thinks there is a 50% chance the number of influenza hospitalizations that week will exceed that number (currently about 18,300) and 50% chance hospitalizations will be lower than it.
There is a community forecast like this for every place (the US as a whole, each state, Washington DC, and Puerto Rico) and every week of the season.
An over-under prediction
Instead of predicting a number yourself, you contribute by predicting whether the count will come in over or under the community forecast, and you commit some points to your prediction. When the week's count is published, the prediction is scored:
- Right, and you get double your points back.
- Wrong, and you get nothing back.
Predictions move the forecast
Your prediction moves the community forecast up (if you chose "over") or down ("under"). "Right" means the count lands at or past the updated forecast; "wrong" means it lands at or short of the prior one. If the count lands in between the prior community forecast and the updated one, then you earn between zero and double your points. Try it:
You commit 10 points (plus a 0.2 fee) saying the count will be over 18,300. The community forecast moves to 21,953. If the count comes in at 21,953 or more, you get 20 back: your points and as many again. At 18,300 or less, nothing. In between, something in between: you break even at about 20,081.
Draw your forecast
Pick a location and draw where you think the weekly count is going. For each week, this produces an over-under prediction in the direction of your line, sized by that week's share of your points (see Points every week below). You can also upload a forecast from your own model in CSV format.
Nothing drawn yet: the community forecast stands where it is.
The 10th and 90th percentiles
For most places, the community forecast also has a 10th percentile and a 90th percentile, the edges of an 80% prediction interval: the community thinks the count will land below the lower line one time in ten, above the higher line one time in ten, and between them eight times in ten. You can predict over or under on these lines too.
The point rewards for 10th and 90th percentile predictions are lopsided. If the 90th percentile is currently 40,900, then the community estimates a 10% chance that the number of admissions is above 40,900 and a 90% chance that it is below. So an "over" prediction is worth 9 times as many reward points as an "under", should it come true.
You commit 10 points (plus a 0.2 fee) saying the 90th percentile should be higher: that the count is more likely to land above 40,900 than the community's one in ten. The line moves to 51,350. If the count comes in at 51,350 or more, you get 100 back, 10× what you committed. At 40,900 or less, nothing. You break even at about 41,861: a rare call pays well, so it only needs to be right a little.
| Prediction | The community's chance of that side | Moves it to | Back if right | As a multiple |
|---|---|---|---|---|
| 90th percentile higher | one in ten | 51,350 | 100 points | 10× |
| 90th percentile lower | nine in ten | 39,879 | 11.11 points | 1.11× |
| median higher | one in two | 21,953 | 20 points | 2× |
| 10th percentile higher | nine in ten | 6,666 | 11.08 points | 1.11× |
| 10th percentile lower | one in ten | 5,178 | 99.92 points | 10× |
Points every week
Every week you sign in you get points to predict with: 1,000 points in each of the season's first 4 weeks, half as many in each of the next 4, and so on. Anything you have not used carries over. Miss a week, or join late, and the next time you sign in you get 80% of each week you missed, on top of that week's own points. For example, if you miss weeks 2 and 3, then in week 4 you get 800 + 800 + 1,000. After that it's back to the usual weekly points. Your points are not meant for one place and one week: by default, each week, everything you have not used (new or carried over) is shared out across every place and every week still open, so that a new forecaster spreads their predictions rather than using them all in the first chart they open.
The soonest weeks (one to four weeks ahead) get the biggest shares of your point commitments at each time. Larger places get bigger shares as well, in proportion to the square root of their population. You can turn off these defaults (training wheels, on your Account page) and allocate your points differently if you choose. For even more flexibility, try the Trader view.
The details
What is being predicted
The count is the number of new hospital admissions with laboratory-confirmed influenza in a week (Sunday to Saturday), for one jurisdiction, as reported by hospitals to the CDC's National Healthcare Safety Network (NHSN). There are 53 jurisdictions: the United States, the 50 states, Washington DC and Puerto Rico. The first count for a week appears the week after it ends, and hospitals keep correcting their reports, so the CDC revises each week's count every Friday for a while, usually upward. We treat a count as final at four weeks from the first official release, and predictions on a week stay open until the day before that, so you can keep predicting while the early counts are being revised. The week is then scored on the final count.
Your points come back to you
Points you commit to a prediction are set aside until its week is scored. Then every prediction on that week is paid out automatically into your pool. What you get back, your weekly allowances and whatever you haven't used yet are all one pool, and you can predict with any of it.
The leaderboards
Your score is what your predictions on scored weeks (weeks whose count is final; see What is being predicted) have paid back, minus the points you committed to them: a prediction that pays back double adds the points you committed to your score, and one that pays back nothing subtracts them, plus every weekly allowance you have received. Points sitting in weeks that haven't been scored count for nothing, for or against, until they are. The leaderboard shows it two ways:
- Finalized counts scored weeks only: weeks whose count is final. This is the score the prize rounds use.
- Provisional (the default) also counts weeks whose count has been published but isn't final yet, as if they had been scored on today's count. It gives an early look, with the caveat that recent counts are preliminary and often revised, usually upward.
There are two prize rounds, and each pays the top five eligible players $1,000, $800, $600, $400 and $200:
- The fall round: the weeks ending Oct 31 through Dec 12, 2026. Its last week is scored around Jan 15, 2027.
- The full-season round: every week ending Oct 31, 2026 through Apr 3, 2027, the fall round's weeks included. Its last week is scored around May 7, 2027.
Each round ranks players by their score counting only that round's weeks, once the last of them is scored. The full rules are in the Terms.
In the trader view
The Trader view exposes participants to the market under the hood (see The science). Aside from additional flexibility in allocating points, the Trader view is the same game, points, and leaderboard, only with different terminology:
| community forecast | the market's prediction: its median, and where a market has them, its 10th and 90th percentiles |
| a prediction, submitting | a trade: moving a percentile's report (the market calls it a quantile) |
| points committed | escrow, plus the trading fee |
| a week is scored | a market is settled |
| points paid back | the position's payout |
| weekly allowance | the faucet's weekly drop of GFLU |