Votara estimate, not an official forecast

How the odds are computed

Votara estimates the chance that a federal vote is accepted. This page explains the method, shows how it would have done on past votes and lists its limits.

Votara does not publish estimates for upcoming votes yet. The first publication needs an editorial sign-off. Here you find the method and a test on past votes – an evaluation of the past, not a prediction.

What is estimated

The chance that an object is accepted on voting day – the official result. Popular initiatives, counter-proposals and mandatory referendums need a majority of both the people and the cantons. The model learns this overall result; it does not compute the two majorities separately.

Switzerland has no open polling data that Votara may republish. The model therefore does not know what voters say in surveys. It only knows the positions of the authorities and the parties and the votes in parliament.

What the model uses

Only information that exists before the vote:

the type of vote (mandatory referendum, optional referendum, popular initiative, counter-proposal);
the recommendation of the Federal Council;
the recommendation of the parliament;
the final votes in the National Council and the Council of States (share of yes votes);
the parties’ recommendations, weighted by their share of the vote at the last National Council election held in an earlier year: the share of parties recommending yes minus the share of parties recommending no.

When a piece of information is missing – for example because parties have not decided yet – the model is told so explicitly. It never fills in a guess. A free vote or a blank-ballot recommendation counts as neither yes nor no. When parties have merged (CVP and BDP into Die Mitte), their earlier share of the vote follows the new party.

How the model learns

A logistic regression, a common statistical method, learns from every federal vote since 1848 how these inputs went together with the results. Recent votes count more: a vote from 20 years ago counts half as much as one from last year.

Settings: regularisation λ = 1, 25 calculation steps. The same data always give the same result.

Backtest: how would the model have done?

Backtest – not a prediction

For every year from 1981 to 2025, the model was trained only on the votes before that year and then estimated that year’s votes. This mimics real use: each estimate only knows the past. 382 votes were estimated. Tie-breaker questions and votes without a result are not included.

Three simple rules serve as a comparison, computed from the same earlier votes with the same weights: the share accepted among votes of the same type; “follow the Federal Council” (the share accepted among earlier votes with the same Federal Council recommendation); “follow the parliament” (the same with the parliament’s recommendation).

Backtest 1981–2025: scores of the model and the comparison rules
Method	Brier score (lower is better)	Log loss (lower is better)	Accuracy
Votara model	0.148(best)	0.464(best)	78.5%
Follow the Federal Council	0.161	0.498	79.1%(best)
Follow the parliament	0.159	0.492	79.1%
Share accepted by type	0.165	0.506	78.3%

Sources: Votara — Votara odds-lite model – probability that a federal vote is accepted (updated on 31.10.2026); Année politique suisse, University of Bern — Swissvotes – database of federal popular votes (data 1848–2027, source checked on 10.11.2026) Checks failing; OFS/BFS — National Council elections since 1919 – party strength and seats (data 1919–2023, source checked on 10.11.2026)

On the Brier score, the model did better than the rule “follow the Federal Council”. Its accuracy, however, was lower: it was on the right side of 50 % for fewer votes, but its probabilities were closer to the outcome on average.

The Brier score is the mean squared gap between the estimate and the outcome (0 = perfect, 0.25 = always saying 50 %). The log loss punishes confident mistakes more. Accuracy is the share of votes where the estimate (above or below 50 %) pointed to the right result.

Calibration: estimated chance and share actually accepted

Horizontal: estimated chance of acceptance. Vertical: share of those votes actually accepted.

Votara estimate, not an official forecast

Share accepted (%)

Perfect calibration
Votara model (backtest)

The chart compares, for 9 groups of estimates (382 votes), the average estimate with the share of votes actually accepted. In 7 groups the actual share was higher than the estimate, in 2 lower or equal. Points on the dotted diagonal would mean perfect calibration.

Show the data as a table

Calibration: estimates in groups and share actually accepted
Estimated chance	Votes	Average estimate	Share accepted
0% to 10%	133	6.3%	9.8%
10% to 20%	26	12.5%	15.4%
20% to 30%	5	25.6%	0%
30% to 40%	11	36.5%	36.4%
40% to 50%	20	46%	70%
50% to 60%	38	55.2%	57.9%
60% to 70%	69	64.9%	69.6%
70% to 80%	52	74.4%	84.6%
80% to 90%	28	83.8%	92.9%

Brier score by decade (lower is better)
Period	Votes	Votara model	Follow the Federal Council	Follow the parliament	Share accepted by type
1981–1989	56	0.159(best)	0.169	0.166	0.168
1990–1999	100	0.144(best)	0.166	0.162	0.169
2000–2009	91	0.139(best)	0.142	0.142	0.149
2010–2019	82	0.133(best)	0.144	0.140	0.153
2020–2025	53	0.183(best)	0.202	0.203	0.197

Scores by type of vote
Type	Votes	Brier score, model	Brier score, follow the Federal Council	Accuracy, model	Accuracy, follow the Federal Council
Counter-proposal	23	0.167	0.254	82.6%	60.9%
Mandatory referendum	71	0.152	0.182	76.1%	78.9%
Optional referendum	125	0.213	0.218	64.8%	68.8%
Popular initiative	163	0.094	0.096	89.6%	89.6%

The fitted model

Fitted on 706 votes from 12 September 1848 to 27 September 2026. A positive number makes acceptance more likely, a negative one less likely. The numbers are log-odds and cannot be compared one by one, because the inputs have different ranges.

Coefficients of model odds-lite-0.1.0
Input	Coefficient
Baseline (optional referendum, all inputs neutral)	-0.81
Mandatory referendum	+0.16
Popular initiative	+0.20
Counter-proposal	-0.19
Federal Council recommendation (+1 yes, −1 no)	+0.27
Federal Council recommendation unknown	-0.08
Parliament recommendation (+1 yes, −1 no)	+0.39
Parliament recommendation unknown	0.00
Yes share in the National Council final vote (minus 0.5)	+0.89
National Council final vote unknown	+0.13
Yes share in the Council of States final vote (minus 0.5)	+0.71
Council of States final vote unknown	+0.09
Vote share of parties recommending yes minus those recommending no	+1.53
Vote share of parties without a known recommendation	+0.05

Known limits

No polls: the model does not see campaigns, turnout or late swings.
Few cases: about ten votes a year, 382 in the backtest. Voting behaviour changes over time.
The 20-year weighting was chosen after seeing the backtest results. This makes the backtest a little too favourable.
The majority of the cantons is not computed separately.
The backtest uses the final recommendations. Before a vote, some parties may not have decided yet.
The vote shares come from the last election. Members switching party during a legislature are not taken into account.

Version and changes

Model version odds-lite-0.1.0.

odds-lite-0.1.0 · 31 October 2026 · First version: 14 inputs, logistic regression, backtest 1981–2025. Not published.

Odds for the three population scenarios

The odds for the three OFS/BFS population scenarios are still in development. Votara will only show them after they have their own backtest.