Adastra: AI Model Predicted The Parliamentary Election Results With Only 12% Of Votes Counted
Artificial intelligence predicted the election results with decimal-point accuracy. Adastra’s AI model demonstrated that data and machine learning can deliver fast, reliable and fully transparent predictions — even in a dynamic environment such as parliamentary elections.
average deviation from the actual results for the 8 largest parties
share of counted votes needed for an accurate prediction
time to the first prediction after CSO data became available
Challenge
Proving the power of AI on data everyone understands
Adastra previously deployed its predictive model during the 2021 parliamentary elections and the 2023 presidential election, where it demonstrated exceptional accuracy.
The goal for the 2025 elections was therefore not experimentation, but proving the robustness and resilience of the system in real time — with a new data format, under media pressure, and with high public attention.
The project served as a real-world validation of a model operating in an environment where there is zero room for error: results are immediately compared with real-world data.
It also became a demonstration of how the same approach can be used in business scenarios where organizations make decisions based on partial or incomplete data — from demand forecasting to manufacturing operations.
Solution
How to turn 12% of the data into a reliable result
Adastra used a predictive model combining election-data analytics, historical trends and machine-learning algorithms.
The original model was developed as part of research by Vojtěch Létal (Blindspot AI) and further enhanced by Adastra — extending its functionality, improving accuracy and integrating it into a modern cloud environment.
The model was trained on historical data from both parliamentary and presidential elections and tested for accuracy and generalization. In the 2023 presidential election, it achieved exceptionally precise results — earning recognition in Forbes as “another winner of the election”.
In 2025, the system processed data from the Czech Statistical Office’s open API, which had changed format shortly before the election. The Adastra team quickly identified the change and adjusted the connection in real time to ensure continuous prediction.
Predictions ran in a scalable cloud environment capable of withstanding high traffic from users and media. The model automatically refined its outputs with every newly counted district and was published in real time by major Czech media — including Seznam, CzechCrunch, Metro, Blesk and Deník N.
Impact
AI proved itself in a real-world test: accuracy, speed and public trust
The first prediction was published at 3:13 p.m., roughly one hour after polling stations closed. Seznam Zprávy published it exclusively, followed by CzechCrunch, Metro, Blesk, Deník N and others.
Newsrooms monitored how the model’s accuracy improved with each new district and compared its outputs with official CSO data.
Predictions started as soon as the first votes were counted, but once the process reached 12%, the estimates stabilized and became reliable enough to publish — with a final average deviation of just 0.34%.
This made the project a reference case for deploying AI in environments requiring transparency, speed and high precision.
Adastra has thus demonstrated that even a small sample of open data is enough to generate a prediction that stands up to public scrutiny.
The know-how gained continues to be applied beyond elections — in projects where the ability to make fast, accurate decisions based on incomplete data is critical, such as:
- demand and market-trend forecasting,
- customer-behavior analytics and retention modelling,
- predictive maintenance in manufacturing.
