Forecast — 2026 German state election in Saxony-Anhalt

Two votes, two roles

Voters in Saxony-Anhalt cast two ballots on the same day. The first vote, the Erststimme (constituency vote), picks a candidate in their constituency: there are 41 of them, and each sends whoever finishes first to the Landtag, by simple plurality in a single round. The second vote, the Zweitstimme (list vote), goes to a party list at the state level.

The two votes do not carry equal weight. It is the second that sets the overall allocation: the total number of seats due to each party is calculated from it alone, proportionally. Direct mandates won on the first vote are then deducted from that total. A party that wins ten constituencies and is entitled to thirty seats overall will receive twenty from its list. AfD is at the extreme end of this mechanism here: 38.5 direct mandates on average, and almost no list seats at all. This is why a model that predicted the constituencies well but the second vote poorly would get the essential part wrong.

The 5% threshold

A party that falls short of 5% of the list vote gets no list seat, whatever its score. Its votes do not simply vanish: they drop out of the calculation, and the allocation runs among the qualifying parties only, which mechanically inflates each of their shares. This is why the story of this election is not the winner but the 5% bar, and why the Greens, polling at 5.15% against a 5% threshold, decide the shape of parliament on their own.

Saxony-Anhalt has no direct-mandate override clause. In Berlin, by contrast, winning just one direct mandate waives the threshold and opens access to the proportional allocation: the same national rule has variants from one Land to another.

Why parliament has no fixed size

The Landtag has a base size of 83 seats. It almost always ends up with more, and that is the least intuitive part of the system.

When a party wins more constituencies than its proportional entitlement allows, it keeps the surplus: these are the overhang seats (Überhangmandate). To restore the balance, the law then allocates the other parties levelling seats (Ausgleichsmandate), until overall proportionality is restored. Parliament therefore grows as a party dominates the constituencies without dominating the list vote.

This is exactly the situation forecast here: AfD is projected to win in 39 of the 41 constituencies with 38.9% of the list vote. The Landtag grows from a base of 83 seats to 91 if the Greens enter, 85 if not. The absolute-majority threshold rises with it, from 43 to 46 seats. This is why AfD can win nearly every constituency and still stay far from a majority. The allocation follows the Hare-Niemeyer method, also known as the largest-remainder method.

Electing the state premier

The Landtag then elects the head of government. Article 65 of the state constitution provides for two rounds requiring an absolute majority of members, then, if no candidate reaches it, a fourteen-day window during which the assembly can vote to dissolve itself. If it does not, a third round is held in which a majority of votes actually cast is enough: abstentions are removed from the count.

This rule is decisive in a parliament without a majority. It allows a state premier (Ministerpräsident) to be elected whom part of the assembly refuses to support, provided that part abstains rather than votes against. The two other Länder voting on September 20 each have their own variant. Mecklenburg-Vorpommern, under Article 42 of its constitution, holds unlimited rounds decided by an absolute majority of members over four weeks: there, an abstention counts as a vote against, so the tolerance scenarios that work in Saxony-Anhalt do not play out the same way. Berlin attaches no dissolution procedure to failure.

A predictive model is not just another poll. A poll surveys a sample and describes opinion at a given moment. A model surveys no one. It gathers what is measurable, published surveys and past election results, and derives from them a distribution of possible outcomes, with their probability.

The question is not who comes first

AfD is projected first in 97.7% of simulations, and wins an average of 38.5 of the 41 direct mandates. That is not the story. The story is the 5% bar: a party that clears it enters the Landtag, a party that misses it gets no seat at all. The Greens are polling at 5.15% and enter in 56.3% of simulations. This single uncertainty shifts six seats, raises the majority threshold by three seats, and cuts AfD's chances of governing alone by five.

And no one has a majority

In 89.7% of draws, no coalition is both majority and politically workable. Article 65(2) of the state constitution allows a state premier to be elected by a majority of votes cast in a third round. The most likely alternative is an AfD absolute majority alone, in 7.2% of draws.

The prior, and why it works

The model does not rely on polls alone. It anchors them in the vote actually cast: the 2021 state election for half its weight, the 2025 federal election for 45%, the 2024 European election for the rest. This prior offsets the polls that got it wrong at the last Saxony-Anhalt state election, where AfD was shown neck and neck with CDU, a race CDU ultimately won.

Our model uses a calculation method based on past results that precisely refines the winner of the election, along with the constituency-level results, to estimate: the seats won in the single-round, single-member constituency vote and the number of seats won through the list vote.

Twenty thousand simulations

A single forecast would give a false sense of certainty. The model therefore replays the election twenty thousand times, varying what is genuinely uncertain: each institute's own error, each party's volatility, and the correlation between them. It then converts each draw into seats under Saxony-Anhalt's electoral law, the Hare-Niemeyer method, with overhang and levelling seats. The Landtag therefore has no fixed size: 91 seats at the median when the Greens enter, 85 if not, and the majority threshold is recalculated for every draw, 46 or 43 depending on the branch.

Two parameters were revised on September 3, after testing the model against three past elections whose result is known: Saxony-Anhalt and Mecklenburg-Vorpommern in 2021, and Berlin in 2023. The historical anchor now gives half its weight to the previous state election instead of a quarter, the European election having proved the least predictive of the three components. And a poll's half-life drops from fourteen to six days: no wave is discarded, but a two-month-old wave now weighs 10% of a same-day wave instead of 37%. All parameters are public, including those we set by judgment. No probability is shown at 0% or 100%: uncertainty is this model's product, never its flaw.

All parameters are public, including those we set by judgment. No probability is shown at 0% or 100%: uncertainty is this model's product, never its flaw.

Sources

Voting intentions: institutes polling on Saxony-Anhalt, aggregated and corrected for bias measured on past elections. Election results: Landeswahlleiter (state returning officer) of Saxony-Anhalt and the Bundeswahlleiterin (federal returning officer), open data under the dl-de/by-2-0 licence. Coalition exclusions: sourced and dated public statements, register available on request. The engine has a reproducible seed, two runs give the same result bit for bit, and it exactly reproduces the 97 seats allocated in 2021. Its 90% interval coverage is measured at 93% across six past elections.

Credits

All rights belong to Le Millénaire SAS, registered at 47 rue de Vivienne, 75002 Paris.

Model designed by
Project director: William Thay
Deputy project director: Pierre Clairé
Political and data analysis: Matthieu Hocque
Data analysis, mathematics and validation: Florian Gérard-Mercier