Jean-Luc Jucker, PhD

SwissRef

SwissRef is a machine learning algorithm that I developed to predict the outcomes of Swiss referendums. It is estimated that on average, SwissRef will correctly predict the results of 70% to 80% of referendums, at best. As of 9 February 2025, SwissRef has correctly predicted the outcomes of 16 out of 23 referendums (over 8 referendum sessions). SwissRef accuracy over time is shown in the graph below.

SwissRef's predictions are less accurate than those of opinion polls. In plain language, SwissRef makes more mistakes than pollsters. Yet it has other advantages. SwissRef can make predictions a long time in advance (up to 2 months before the voting day), and its predictions do not change over time. SwissRef is also more cost-efficient, since it does not require any survey or poll to be conducted to make predictions. Instead it uses exclusively machine learning methods, in particular, extensive feature engineering.

As such, SwissRef can be used as a starting point or 'baseline' to predict the result of votations. It provides estimates as to how people are expected to vote in the future, provided that they behave as in past votations.

SwissRef is experimental and currently under development, and has important limitations. It should therefore be used with caution.

You can read more about SwissRef and why and how I developed it in the FAQ, which was written for non-specialists.

SwissRef Accuracy Over Referendum Sessions (cumulative)

How to read: up to and including the 5th session, SwissRef accuracy was just above 80%. Up to and including the 7th session, SwissRef accuracy was just under 70%. Etc. Number of referendums up to and including the most recent session = 23. The baseline shows the proportion of the most frequent outcome, i.e., rejected referendums.

Forthcoming referendums

28 September 2025 (session 9)

Status: Pending

Past referendums

9 February 2025 (session 8)

(677) Popular Initiative for Environmental Responsibility (Environment)

Predicted majority vote: NO

Actual majority vote: NO

Status: CORRECT

Prediction date: 9 January 2025 (31 days before the voting day), see here. Also see notes below for more information.

24 November 2024 (session 7)

(673) Expansion programme for the national highways (National Highways)

Predicted majority vote: YES

Actual majority vote: NO

Status: INCORRECT

(674) Tenancy law: subletting (Subletting)

Predicted majority vote: YES

Actual majority vote: NO

Status: INCORRECT

(675) Tenancy law: termination for personal use (Personal Use)

Predicted majority vote: YES

Actual majority vote: NO

Status: INCORRECT

(676) Health insurance: standardised benefit financing (Health Insurance)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

Prediction date: 21 October 2024 (34 days before the voting day), see here. Also see notes below for more information.

22 September 2024 (session 6)

(671) Biodiversity Popular Initiative (Biodiversity)

Predicted majority vote: NO

Actual majority vote: NO

Status: CORRECT

(672) Occupational Pension Reform (Occupational Pension)

Predicted majority vote: YES

Actual majority vote: NO

Status: INCORRECT

Prediction date: 28 August 2024 (25 days before the voting day), see here. Also see notes below for more information.

9 June 2024 (session 5)

(667) Popular Initiative for Health Insurance Premiums Relief (Max 10%)

Predicted majority vote: NO

Actual majority vote: NO

Status: CORRECT

(668) Popular Initiative for Reducing Health Insurance Costs (Reducing Costs)

Predicted majority vote: NO

Actual majority vote: NO

Status: CORRECT

(669) Popular Initiative for Physical Integrity (Vaccination)

Predicted majority vote: NO

Actual majority vote: NO

Status: CORRECT

(670) Law on Electricity (Electricity)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

Prediction date: 6 May 2024 (34 days before the voting day), see here. Also see notes below for more information.

3 March 2024 (session 4)

(665) Popular Initiative for a 13th Pension (13th)

Predicted majority vote: NO

Actual majority vote: YES

Status: INCORRECT

(666) Popular Initiative for an Increased Retirement Age (Retirement)

Predicted majority vote: NO

Actual majority vote: NO

Status: CORRECT

Prediction date: 28 January 2024 (35 days before the voting day), see here. Also see notes below for more information.

18 June 2023 (session 3)

(662) Taxation of Large Companies (OECD/G20)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

(663) Law on Protection of Climate and Energy Security (Climate)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

(664) Amendment to the Law on COVID-19 (COVID)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

Prediction date: 8 May 2023 (41 days before the voting day), see here. Also see notes below for more information.

25 September 2022 (session 2)

(658) Popular Initiative on Intensive Animal Farming (Intensive Farming)

Predicted majority vote: NO

Actual majority vote: NO

Status: CORRECT

(659) Funding of Old-Age Insurance through VAT increase (AVS 21a)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

(660) Amendment to the Law on Old-Age Insurance (AVS 21b)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

(661) Amendment to the Law on Withholding Tax (Withholding Tax)

Predicted majority vote: YES

Actual majority vote: NO

Status: INCORRECT

Prediction date: 28 July 2022 (59 days before the voting day), see here. Also see notes below for more information.

15 May 2022 (session 1)

(655) Amendment to the Law on Cinema (Cinema)

Predicted majority vote: NO

Actual majority vote: YES

Status: INCORRECT

(656) Amendment to the Law on Transplantation (Transplantation)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

(657) EU Regulation on Border Management (Frontex)

Predicted majority vote: YES

Actual majority vote: YES

Status: CORRECT

Prediction date: 18 April 2022 (27 days before the voting day), see here. Also see notes below for more information.

Notes

Referendum of 9 February 2025 (past)

Predicted outcome: The probability associated with the predicted outcome (classification task) is 56.31% (Environment). Furthermore, the predicted percentage of Yes votes (regression task) is 35.50% (Environment), with a Mean Absolute Error (MAE) of 8.67. Model version: 2.1. Actual outcome: Environment: rejected (30.25% Yes).

Referendums of 24 November 2024 (past)

Predicted outcomes: The probabilities associated with the predicted outcomes (classification task) are 78.98% (National Highways), 52.45% (Subletting), 62.89% (Personal Use), and 81.63% (Health Insurance). Furthermore, the predicted percentage of Yes votes (regression task) are 54.16% (National Highways), 51.20% (Subletting), 53.62% (Personal Use), and 54.92% (Health Insurance), with a Mean Absolute Error (MAE) of 8.34. Model version: 2.1. Actual outcomes: National Highways: rejected (47.29% Yes); Subletting: rejected (48.41% Yes); Personal Use: rejected (46.16% Yes); Health Insurance: accepted (53.31% Yes).

Referendums of 22 September 2024 (past)

Predicted outcomes: The probabilities associated with the predicted outcomes (classification task) are 79.51% (Biodiversity), and 73.86% (Occupational Pension). Furthermore, the predicted percentage of Yes votes (regression task) are 41.34% (Biodiversity), and 58.86% (Occupational Pension), with a Mean Absolute Error (MAE) of 8.38. Model version: 2.1. Actual outcomes: Biodiversity: rejected (36.96% Yes); Occupational Pension: rejected (32.88% Yes).

Referendums of 9 June 2024 (past)

Predicted outcomes: The probabilities associated with the predicted outcomes (classification task) are 84.41% (Max 10%), 87.23% (Reducing Costs), 77.88% (Vaccination), and 54.61% (Electricity). Furthermore, the predicted percentage of Yes votes (regression task) are 45.07% (Max 10%), 40.19% (Reducing Costs), 44.55% (Vaccination), and 70.82% (Electricity), with a Mean Absolute Error (MAE) of 10.04. Model version: 2.1. Actual outcomes: Max 10%: rejected (44.52% Yes); Reducing Costs: rejected (37.23% Yes); Vaccination: rejected (26.26% Yes); Electricity: accepted (68.71% Yes).

Referendums of 3 March 2024 (past)

Predicted outcomes: The probabilities associated with the predicted outcomes (classification task) are 86.59% (13th), and 81.39% (Retirement). Furthermore, the predicted percentage of Yes votes (regression task) are 35.50% (13th), and 34.46% (Retirement), with a Mean Absolute Error (MAE) of 9.12. Model version: 2.1. Actual outcomes: 13th: accepted (58.25% Yes); Retirement: rejected (25.25% Yes).

Referendums of 18 June 2023 (past)

Predicted outcomes: The probabilities associated with the predicted outcomes (classification task) are 53.29% (OECD/G20), 50.70% (Climate), and 54.26% (COVID). Furthermore, the predicted percentage of Yes votes (regression task) are 52.11% (OECD/G20), 51.51% (Climate), and 51.52% (COVID), with a Mean Absolute Error (MAE) of 9.55. Model version: 2.0. Actual outcomes: OECD/G20: accepted (78.45% Yes) ; Climate: accepted (59.06% Yes) ; COVID: accepted (61.93% Yes).

Referendums of 25 September 2022 (past)

Please note that AVS 21a and AVS21b are tied to each other: if one is rejected by the voters, the other one is automatically rejected as well.

Predicted outcomes: The probabilities associated with the predicted outcomes (classification task) are 55.00% (Intensive Farming), 53.71% (AVS 21a), 53.58% (AVS 21b), and 54.07% (Withholding Tax). Furthermore, the predicted percentage of Yes votes (regression task) are 32.78% (Intensive Farming), 58.33% (AVS 21a), 53.38% (AVS 21b), and 55.93% (Withholding Tax), with a Mean Absolute Error (MAE) of 9.01. Model version: 1.1. Actual outcomes: Intensive Farming: rejected (37.13% Yes); AVS 21a: accepted (55.07% Yes); AVS21b: accepted (50.54% Yes); Withholding Tax: rejected (47.99% Yes).

Referendums of 15 May 2022 (past)

Predicted outcomes: The probabilities associated with the predicted outcomes (classification task) are 57.20% (Cinema), 89.66% (Transplantation), and 85.64% (Frontex). Furthermore, the predicted percentage of Yes votes (regression task) are 42.02% (Cinema), 56.00% (Transplantation), and 53.66% (Frontex), with a Mean Absolute Error (MAE) of 8.23. Model version: 1.1. Actual outcomes: Cinema: accepted (58.42% Yes); Transplantation: accepted (60.20% Yes); Frontex: accepted (71.48% Yes).

FAQ

What is SwissRef?

SwissRef is a tool to predict the outcomes of Swiss referendums. In other words, to predict whether Swiss voters will accept or reject the object of a referendum. SwissRef also allows to estimate the percentage of voters who will accept or reject a referendum.

How accurate is SwissRef?

It is estimated that SwissRef will make correct predictions in 70% to 80% of cases, at best. In other words, that it will correctly predict the outcome of 7-8 out of 10 referendums on average, at best. This estimation is based on machine learning estimation methods applied to the outcomes of previous referendums, in particular, a method called 'cross-validation'.

Furthermore, it is estimated that SwissRef will correctly predict the percentage of yes or no votes within a range of -9 to +9 percentage points, at least. For example, if SwissRef predicts that a referendum will obtain 61% of yes votes, the actual percentage could fall anywhere between 52% and 70% of yes votes, at least.

Please note that since SwissRef has just been launched, it is not possible to compute its actual or 'real-world' accuracy yet. For example, it is possible that SwissRef starts with a low accuracy, to then gain more accuracy, or vice-versa. The accuracy figures reported above only mean that over time and with more predicted outcomes, SwissRef should tend to an accuracy of 80% and an error of -9 to +9 at best, on average.

As of 24 November 2024, SwissRef has correctly predicted the outcome of 15 out of 22 referendums.

How long in advance can SwissRef make predictions?

In principle, SwissRef can make predictions several months before the voting day. In practice, I aim to generate predictions at least 1 month before the voting day, when the campaign starts. I aim to generate predictions before any opinion poll has been conducted.

Likewise, once the predictions have been generated and published here and on Twitter (the same day), I aim to not change them, and have done so up to today. If this were the case, the change would be clearly highlighted here and re-published on Twitter with an explanation.

What are the advantages and disadvantages of SwissRef?

The main disadvantage of SwissRef is that it is less accurate than other predictive tools. In other words, it makes more mistakes than other predictive tools.

The main advantages of SwissRef are that it can make predictions a long time before other predictive tools (up to 1 month before any of them), that its predictions do not change over time, and that it is highly cost-efficient (no opinion polls are needed).

How can SwissRef be used?

Since SwissRef is less accurate than other predictive tools, it cannot be used to predict the outcome of referendums with a high level of reliability. For more accurate and real-time predictions, I recommend using gfs.bern and Predikon.

The best way to use SwissRef is to see it as a ‘reference point’ or ‘baseline’ of how people are expected to vote. In other words, SwissRef provides a general indication of how people will vote in the future, based on how they voted in the past on similar referendums.

What are the differences between SwissRef and other predictive tools?

Other predictive tools in Switzerland include those developed by gfs.bern, and Predikon.

GFS Bern uses traditional opinion polls to make predictions at regular intervals, with the last opinion poll usually conducted 2 weeks before the voting day. Predikon uses machine learning (and more recently, informal opinion polls) to make real-time predictions on the voting day.

In contrast, SwissRef makes predictions without conducting any opinion poll (contrary to GFS Bern), and this, a long time in advance (contrary to Predikon). Although it is less accurate than both GFS Bern and Predikon, SwissRef has other advantages and can therefore be used as a complementary tool.

How does SwissRef make its predictions?

SwissRef uses machine learning methods to make predictions. In particular, SwissRef uses extensive feature engineering. A range of models (Logistic, Ridge, k-NN, random forests, etc.) are fitted and compared before selection. Hyperparameters are tuned using grid search, and model performance is evaluated using cross-validation. SwissRef does not use any survey or opinion poll data to make predictions.

Who created SwissRef?

I am the creator and sole contributor of SwissRef. I developed SwissRef using public, open-source data only. I am solely responsible for any mistake or error that SwissRef might contain. SwissRef is not funded by any organization (public or non public) or individual, I have no competing interest to declare, and I am not a member of any Swiss political party or political organization. That said, I reserve the right to use SwissRef for any purpose I see fit in the future, including for commercial purposes such as consulting, in which case this would be stated it here.

Why was SwissRef created?

Predicting the outcomes of votations is difficult. On the one hand, machine learning projects such as Predikon are highly accurate, but they cannot make predictions before the voting day, when the first results from the municipalities are known. In other words, these predictions are accurate, but not very useful, since it is too late for the government, political parties, or lobbying groups to do anything about the outcome (such as inform their campaign, for example). On the other hand, opinion polls such as those of gfs.bern are available several weeks before the voting day, yet they are not as accurate as Predikon, and sometimes, incorrect. It also happens that pollsters only present trends, but do not make an actual prediction. Furthermore, opinion polls are costly: a representative sample of the population must be surveyed (often, several times), meaning time and resources to invest for pollsters, and spending public money.

Importantly, a good prediction should be both accurate and useful. In the field of votations, this means both (a) predicting the outcome of a votation with a high accuracy (as Predikon does), and (b) a long time in advance (as opinion polls do). A good predicting system should also be (c) as cost-efficient as possible.

In other words, predicting an outcome accurately, usefully, and cost-efficiently is the holy grail of machine learning! But again, this is difficult, and when I started developing SwissRef (while a learner at the EPFL), my goal was slightly less ambitious. It was to fill a gap between existing predictive tools, as illustrated in the Venn diagram below. Other machine learning projects such as Predikon are accurate and cost-efficient, but not very useful. Opinion polls are more useful, but slightly less accurate, and much less cost-efficient, than Predikon. In contrast, SwissRef is less accurate than both Predikon and opinion polls, yet it is more useful than Predikon, and more cost-efficient than opinion polls. It can therefore be used as a complementary tool, whose main advantages are that it generates predictions a long time before competitors, and that its predictions are stable over time. Also note that I have not given up on the holy grail, and that I constantly develop and improve SwissRef!

When was SwissRef launched?

SwissRef was launched on 18 April 2022, on this website. Its first predictions were made on the same day (18 April 2022), and concerned the referendums of 15 May 2022.

Are the data and code available?

All the data used to develop SwissRef are public open-access data, and can be downloaded on various Swiss official websites, for example, here, here, and here. In contrast, I have not made the Python code available as of today.

How often is this page updated?

This page is updated every time that I generate new predictions, which is normally 4 times per year. After each votation day, I also report the actual results, which is also, normally, 4 times per year. I sometimes make cosmetic changes or add notes and precisions. Of course, I do not make any changes to the predictions once these have been generated. Predictions are published on my Twitter account, meaning that they are time-stamped.

What could go wrong?

This is a small-scale project, which I developed in my free time, and predicting the outcomes of referendums is especially difficult . The most likely thing that could go wrong is that SwissRef generates a higher number of incorrect predictions than expected. In which case, SwissRef should not harm anyone, except myself :-) Likewise, due to its small scale and limited resources, SwissRef is unlikely to have any negative societal impact. At worst, people might not understand what SwissRef is trying to achieve, or overestimate its accuracy. To mitigate this risk, I wrote this FAQ and clearly acknowledged the many limitations of SwissRef. I also always mention the limitations of SwissRef and redirect to the current page when I talk about it on social media. Finally, I am as transparent as possible concerning my potential conflicts of interest (see Who created SwissRef?), the nature of this project, and changes made to this page.

That said, more generally, what is the value of having predictive models for democratic decisions? Who benefits from having these predictions made? Do these models support citizens in decision-making, and ultimately, the common good? These are important questions, and I can only give incomplete answers here. Here are a few examples that applies specifically to Switzerland.

First, a good predictive model might help political parties and activists to spend their limited resources and money more efficiently. For example, if a model with a proven high accuracy (which is not the case of my model) predicts that an initiative has an extremely low chance to be accepted by voters, they might use that insight to redirect their resources to another initiative that is important to them, but which has a higher chance of being accepted by voters. Second, a good predictive model might help to save public money. For example, public TV could focus on informing people primarily (although not only, of course) on those referendums where the result is expected to be tight, suggesting that the subject is also more complex or divisive, and that voters therefore need more impartial information about it. Finally, a good predictive model might motivate citizens to vote, especially when the predicted outcome is extremely tight. In other words, a good model might increase the participation of citizens to democratic processes when this matters most.

As a final note, I am aware that since my model has a relatively low accuracy, it cannot directly be used regarding the above examples (except concerning initiatives, perhaps). I am also aware that one could cite other examples where predictive models, on the contrary, undermine democratic processes. My point here is only that predictive models also have potential for the common good, and that they do not necessarily harm.

I thank Luc Henry for suggesting to write this section of the FAQ.

Page last updated: 3 March 2025. Adding result of referendum of 9 February 2025; updated Introduction and Accuracy plot.

Previous changes to this page:

9 January 2025. Adding prediction for referendum of 9 February 2025.

3 December 2024. Adding results of referendums of 24 November 2024; edits in Introduction (new estimation of SwissRef accuracy at 70-80%) and FAQ (idem); added a graph showing SwissRef accuracy over referendum sessions.

21 October 2024. Adding predictions for referendums of 24 November 2024; harmonization of abbreviations.

29 September 2024. Adding results of referendums of 22 September 2024; two edits in Introduction.

28 August 2024. Adding predictions for referendums of 22 September 2024.

17 June 2024. Adding results of referendums of 9 June 2024; two edits in Introduction.

6 May 2024. Adding predictions for referendums of 9 June 2024; adding reference numbers for all referendums; one edit in FAQ.

4 March 2023. Adding results of referendums of 3 March 2024; two edits in Introduction.

28 January 2024. Adding predictions for referendums of 3 March 2024.

27 July 2023. Adding results of referendums of 18 June 2023; two edits in Introduction.

8 May 2023. Adding predictions for referendums of 18 June 2023; one edit in FAQ.

25 October 2022. Adding results of referendums of 25 September 2022; minor edits in FAQ and Introduction.

28 July 2022. Adding predictions for referendums of 25 September 2022; major edits in FAQ (added sections); minor edits in Introduction.

Google Sites

Report abuse