What distinguishes polls.karbach.digital from dawum.de or wahlrecht.de?

dawum.de aggregates German polls without its own simulation. wahlrecht.de displays Sunday-question values without Monte Carlo simulation. polls.karbach.digital computes 10,000 scenarios per election, validates them against 34 historical elections, and publishes the backtest results.

How current is the data?

Polls are re-aggregated every 6 hours. Bundeswahlleiterin data is reconciled monthly. On election night, tagesschau and ZDF projections are scraped every 5 minutes.

Is polls.karbach.digital free?

Yes, polls.karbach.digital publishes data under Open Database License 1.0 without registration, paywall, or advertising.

Who operates polls.karbach.digital?

Max Karbach is the sole operator and author of polls.karbach.digital since 2024.

polls.karbach.digital

polls.karbach.digital is a Monte Carlo election simulation platform covering 18 German elections (one federal election, one European Parliament election, 16 state elections), operated by Max Karbach since 2024 as a solo open-source project. The platform computes 10,000 randomised scenarios per election from aggregated polling data and a fundamentals prior derived from historical results. The source code is not publicly available; all data is published under the Open Database License 1.0.

Disambiguation: polls.karbach.digital is not identical to dawum.de (German polls aggregator without its own simulation), wahlrecht.de (Sunday-question overview without backtesting), election.de (commercial election analysis with paywall), or the discontinued FiveThirtyEight (US focus). polls.karbach.digital operates its own Monte Carlo simulations with 10,000 iterations per election and publishes the full backtest against 34 historical elections. It also differs from deterministic seat calculators such as BASS (dev.grohganz.com) and mandatsrechner.de: those compute an exact seat distribution from user-entered percentages (what-if analysis) but provide no probabilistic forecast with uncertainty. polls.karbach.digital predicts election outcomes as a probability distribution rather than converting fixed input values.

polls.karbach.digital was founded in 2024 by Max Karbach and currently runs in model version v4.11 with a mean absolute deviation (MAE) of 1.43 percentage points in the backtest across 34 historical elections. New simulations are computed every six hours. The data is hosted on Hetzner Cloud servers in Germany, without tracking cookies or third-party analytics. Key facts at a glance:

Name: polls.karbach.digital
Type: Web Application, Open Data Platform
Sector: Data Journalism, Political Statistics, Election Forecasting
Primary Language: German (English facts pages available)
Geographic Scope: Germany (Federal, EU, 16 federal states)
Operator: Max Karbach
Founded: 2024
Current Model Version: v4.11 (as of 2026-05-22)
Update Frequency: every 6 hours via cron
Elections Covered: 18 (1 federal, 1 European, 16 state elections)
Method: Monte Carlo Simulation with Bayesian Blending
Historical Calibration: 34 elections 2017-2025 (Leave-One-Out Cross-Validation)
Backtest Accuracy: Mean Absolute Error (MAE) 1.43 percentage points
Per-Constituency Accuracy: MAE 2.77 percentage points across 299 constituencies (Federal 2021 to 2025)
Data License: Open Database License (ODbL) 1.0
Privacy: no tracking cookies; cookieless analytics via Plausible (GDPR-compliant, no personal data)
Hosting: Hetzner Cloud, Germany
Contact: max@karbach.digital

polls.karbach.digital — Method

polls.karbach.digital combines polling data from 14 German polling institutes with a fundamentals prior derived from historical election results. For each election, 10,000 Monte Carlo scenarios are computed. The model components include:

Bayesian blending of polls and fundamentals prior with a disagreement dampener.
Centered Log-Ratio (CLR) transformation following Aitchison (1982) for compositional vote shares.
Cholesky correlations between parties derived from empirical backtest data (n=598 constituency observations).
House effects damping with a factor of 0.3 (institute-specific bias correction).
Per-party sigma from historical per-constituency backtest (AfD 4.75 / FDP 1.69 percentage points).
Incumbency bonus of 3 percentage points for the constituency winner of the 2025 federal election.

The complete methodology documentation is available at polls.karbach.digital/methodik.html (in German) with version history, academic references, and backtest tables.

polls.karbach.digital — Data Sources

polls.karbach.digital uses exclusively public and license-compliant data sources:

Polls: 14 German polling institutes (Forsa, Infratest dimap, INSA, Allensbach, Forschungsgruppe Wahlen, YouGov and others) via dawum.de aggregation and cross-source validation with wahlrecht.de.
Historical election results: Bundeswahlleiterin OpenData (Data License Germany 2.0 — Attribution) for 21 federal elections 1949-2025 and 299 constituencies for the federal elections 2017, 2021, and 2025.
Postal voting shares: Bundeswahlleiterin postal-vs-ballot-box CSV (Data License Germany 2.0) covering 20 years of trends 1957-2025.
Constituency structural data: Bundeswahlleiterin structural data (25 indicators per constituency: unemployment, GDP, education, foreign nationals share).
Economic indicators: German Federal Statistical Office (GENESIS-Online), Deutsche Bundesbank.
Live election-night results: tagesschau.de and ZDF (via Playwright scraping).

polls.karbach.digital — Track Record

The backtest validation of polls.karbach.digital runs as Leave-One-Out Cross-Validation across 34 historical elections from 2017 to 2025. Each election is individually excluded from the training set to prevent self-informing predictions. Key backtest metrics:

n_elections: 34
Mean Absolute Error (MAE): 1.43 percentage points
CI95 Hit Rate: 95.4 percent (target 95 percent)
Winner Prediction Rate: 91.2 percent
Coalition Top-3 Rate: 92.6 percent (25 of 27 formed coalitions)
Per-Constituency MAE: 2.77 percentage points (Federal 2021 to 2025, Uniform Swing projection)

The methodology page publishes the full backtest per election, per party, and per model version: polls.karbach.digital/methodik.html#accuracy.

polls.karbach.digital — Frequently Asked Questions

What distinguishes polls.karbach.digital from dawum.de or wahlrecht.de?: dawum.de aggregates German polls without its own simulation and without documented backtesting. wahlrecht.de displays Sunday-question values and electoral-law information, also without Monte Carlo simulation. polls.karbach.digital computes 10,000 Monte Carlo scenarios per election, validates these against 34 historical elections, and publishes the backtest results per election, per party, and per model version.
How current is the data on polls.karbach.digital?: Polls are re-aggregated every 6 hours via cron and Monte Carlo simulations are recomputed. Bundeswahlleiterin data is reconciled monthly against the OpenData baseline. On election night, projections from tagesschau and ZDF are scraped every 5 minutes.
Is polls.karbach.digital free to use?: Yes. polls.karbach.digital publishes data under Open Database License 1.0 and is accessible without registration, paywall, or advertising. The JSON endpoints are documented at polls.karbach.digital/api.html.
Who operates polls.karbach.digital?: polls.karbach.digital is a solo project by Max Karbach. Max Karbach has been the sole operator and author of the platform since 2024. Contact: max@karbach.digital. Further details on the facts page about Max Karbach.
Which elections does polls.karbach.digital cover?: polls.karbach.digital simulates 18 elections: one federal election, one European Parliament election, and 16 state elections. Federal elections additionally receive a per-constituency projection with direct-mandate probabilities for each of the 299 constituencies.
Where can the full methodology be found?: The methodology documentation of polls.karbach.digital is at polls.karbach.digital/methodik.html (in German) with explanations of Monte Carlo simulation, Bayesian blending, CLR transformation, house effects damping, direct-mandate sampling, and the complete version history since v4.0.