Experiment
Can a forecast beat "same as this week"?
We're testing whether a forecast can do better than simply assuming the next two weeks look like this week. We write each forecast down before the real number is published, and score it in the open. Until it clearly does better, we don't show forecasts on the main page.
The score so far
Interim, not a verdict. The verdict comes after week 20 of 2027.
Loading the track record.
Latest results, country by country
For each country, the newest forecast whose real number has now been published. A forecast is only shown here once its week has passed, so this page is a record and not an outlook. "Points" are percentage points of tests that were positive: going from 10% to 11% is one point.
Forecast against what happened
Loading.
The "likely falling" flag
Loading.
Every scored forecast
Show all of them, including the bad misses
The rules we set in advance
We wrote these rules on 19 November 2026, before the first forecast was made, and we have not changed them. If we ever have to, the change will be listed here with its date and the reason.
What is being forecast
The share of positive flu tests, among tests done by family doctors in a country, two weeks ahead (ECDC weekly data). We compare the forecast with "same as this week", which simply repeats this week's number. We use the first number ECDC publishes for a week, not later corrections, and we leave out any week with fewer than 50 tests, because that number is too noisy to judge anything.
Which countries, and when
Thirteen countries, fixed for the season: Austria, Cyprus, Denmark, France, Germany, Greece, Hungary, Ireland, Italy, Luxembourg, Norway, Poland and Spain. They are the ones where the forecasts were already tried on past winters and the data is long enough. A country with too few tests in a given week simply gets no forecast that week; it is not counted as a miss. Each time ECDC publishes a new latest week, we write down the forecasts that day. Forecasts made from weeks before 2026 week 46 are kept in the log but not counted, and we count forecasts for target weeks up to week 20 of 2027.
What the forecasts are
The model is Timestral, a small public forecasting model released in October 2026. We use it as it comes, with no tuning to flu, in a small version (5M) and a larger one (123M). Each forecast comes with a "likely range", the range the model expects the real number to fall into 8 times out of 10. We also write down what "same as this week" would have said.
What counts as doing better
The small model has to beat "same as this week" on all of these, over the whole season:
- Its typical miss is at least 10% smaller.
- That gain is bigger than chance, which we check by re-sampling the countries 1,000 times.
- It also scores better on a measure that rewards good ranges, not just the middle guess.
- Its likely range holds the real number between 7 and 9 times out of 10.
- No group of countries (more tests or fewer tests) does more than 5% worse than "same as this week".
We also need at least 100 scored forecasts, at least 10 weeks of them, and at least 8 countries with 5 or more each. Without that, the answer is "not enough data", which counts as not passed.
The halfway check
After week 5 of 2027 we look once. The check can only stop the experiment if a model is clearly doing harm: at least 50 scored forecasts and a typical miss more than 10% bigger, or a likely range that holds the real number less than 55% of the time. It can never declare a pass.
The "likely falling" flag
Separate from the above. The model calls "likely falling" when its two-week forecast is at least 5 points below this week. The call is right if the real number is then more than 5 points lower. To pass, it needs at least 30 calls, at least 65% of them right, and a margin above 50% that is bigger than chance. We never show a "rising" or "steady" flag.
No editing the past
Every forecast is saved before its real number exists, in a log that is only ever added to. A mistake is fixed by a new dated entry, never by changing the old one. A week we missed is recorded as a gap, and we do not fill it in later.
What this test cannot tell us
It covers one season and about 13 countries, so even a real improvement could fail to show. Flu is noisy: a few tenths of a point can decide the result. In one of the three past winters we tried, the model gave no gain at all, so a fail is a likely and honest answer.
What happens at the end
If a model passes (the small one if both do), we will add a clearly labelled "Experimental" two-week outlook for the countries it was tested in, showing the range and what "same as this week" says. If not, nothing changes on the main page and we report the result here.