How the forecast works

The data, the model and every test behind the numbers, including what it gets wrong.

In one paragraph

Douze Points plays the Eurovision Song Contest 2027 thousands of times. In every run, each of the 36 countries' juries and publics ranks the songs and hands out 12, 10, 8 down to 1 point, through both semi-finals and the grand final. How a country ranks a song depends on three things: how strong that song is expected to be, the voting habits between the two countries, and luck on the night. The share of runs a country wins becomes its percentage on the board.

The data

  • Every vote since 1957. individual point awards from grand finals and semi-finals, 1957 to 2026, from the open Eurovision dataset (Spijkervet et al., built from Eurovisionworld), extended with Eurovisionworld scoreboards for the 2023 semi-finals and 2024–2026.
  • Juries and public separately since 2016, when each country began giving two sets of points: separate ballots.
  • Checked against the official results. Every grand-final total from 2016 to 2026, jury and public, matches the official scoreboards exactly.
  • Corrections we made. The source listed 2023 qualifiers twice (a semi-final row and a final row), which we merged. It also lists every 1956 song as second, because only the winner was announced; we added Switzerland's win by hand.
  • Betting odds for the final week, 2017–2026, from the Eurovisionworld odds archive, used only to test the model and, from 2027, for countries whose songs are chosen.
  • Song lyrics in the dataset are copyrighted and are not used or published.

The model

For each voter (a country's jury, or its public) and each song, the model draws a score and the voter ranks the songs by it:

score = song strength + friendship + night shock + personal taste
  • Song strength. Before songs exist, a country's average result over its last ten contests, every year counting the same (the share of the winner's score it reached, with semi-final exits counted low). Once a country's song is chosen and the betting market prices it, its strength comes from those odds instead.
  • Friendship. How many more points this voter has given this country than it gives an average entry, every contest since 2004, newer contests weighted more and small samples shrunk toward zero. It is measured separately for juries and the public, because they love different neighbors.
  • Night shock. How well the song lands on the night, shared by every voter. Sixty percent of it is shared between juries and public, so a song can be a jury darling and a public flop. It is large before songs are known and shrinks as songs arrive.
  • Personal taste. A random term for each voter, so two neighbors don't always vote alike.

Until the official allocation draw, each run draws the two semi-finals the way the EBU does: countries sit in pots of voting neighbors (built here from the same voting habits), each pot is split across both semis, and Armenia and Azerbaijan are kept apart. Once the real draw is announced, the official line-ups replace this. Each run then plays the semis, sends ten from each to the final with the Big Five and the host, and plays the final. We run of these.

How we tested it

1. Before songs exist (15 contests)

We replayed every final from 2010 to 2026 the way the model would have seen it a year earlier, using only past results and voting habits. The live settings come from this test.

SettingRanks the field (rank correlation)Winner log loss (lower is better)
Pure chance0.003.25
Last six contests, recent years weighted heavily0.223.28
Last ten contests, weighted equally (live)0.293.20

History alone ranks the field clearly better than chance but barely beats chance at naming the winner. That is why no country passes 6% today.

2. As songs arrive

On the nine finals from 2017 to 2026, we revealed a share of the songs through their betting odds and kept history for the rest. Average chance given to the eventual winner:

Songs knownChance given to the winnerRank correlation

The test used final-week odds, which know more than odds set right after a national final, so real gains mid-season will be smaller.

3. On the night (9 finals)

With every song known and the closing odds as song strength, adding voting habits made the forecast sharper than the bookmakers alone:

YearWinnerModel's chanceBookmakers' chance

In 2017 the closing odds had Bulgaria and Portugal almost level (2.75 and 2.91). Pre-show odds that day had Portugal ahead. So the bookmakers' favorite called 4 of 9 winners by closing odds, or 5 by pre-show odds. The model called 5.

4. Who qualifies (20 semi-finals)

For every semi-final from 2016 to 2026 we knew the line-up and who voted, and asked the pre-season model who would qualify. Voting habits turned out to matter about twice as much in a semi-final as in the final, and the night drifts less from expectations. Settings chosen on nine years and checked on the tenth beat the old ones in 6 of 10 years.

Semi-final settingLog lossBrier scoreQualifiers named
Coin flip (everyone the same)0.6620.23462%
Same as the final (before)0.6470.22967%
Semi-final settings (now)0.6400.22468%

Lower is better for log loss and Brier score. "Qualifiers named" is how many of the model's ten favorites in each semi went through, over 319 entries.

5. What we tried that did not help

Ideas that made no difference beyond chance, so the live model leaves them out:

  • Separate jury and public strength. Since 2016 juries and the public score separately, so a country can be strong with one and weak with the other. Before songs exist, splitting them changed the winner log loss by less than 0.01.
  • Faster or slower fading of voting habits. Counting old friendships less, or not discounting them at all, made no clear difference.
  • A qualifying record. Adding how often a country got out of its recent semi-finals made qualifier forecasts slightly worse: recent results already carry that.

What it gets wrong

  • It was tuned on the same contests it was tested on. Nine to fifteen contests is a small sample. 2027 is its first real test.
  • It cannot hear songs. Before odds exist, it knows nothing about the music, the staging or the performer.
  • Voting rules change. Semi-final voting has changed several times (no juries in 2023–2025, back in 2026); the model uses one set of semi-final settings for all of them.
  • New and returning countries have little or no history. Canada is treated as an average newcomer.
  • Diasporas and politics move. Friendships are weighted toward recent contests, but sudden shifts take time to show.

Open data

Everything on the site is free to reuse with credit to Douze Points (douze.valevlaube.com). The files update with the forecast.

Changes