Qminers Quant Hackathon 2025: The Most Thrilling Edition Yet!

The fourth edition of the Qminers Quant Hackathon is behind us, and the bar has moved yet again. This year, around 250 participants in roughly 120 teams joined the competition, a major leap from last year. Twenty top teams advanced to the final round, and three of them split prize money totaling CZK 100,000. The competition was tougher, the data more complex, and the atmosphere excellent, exactly what we want from Quant Hackathon.

What is the Qminers Quant Hackathon and why do we run it? It is a student competition built on the same principles as algorithmic trading: spotting the mismatch between “price” and reality, estimating probabilities well, and managing risk properly. Instead of market data, participants work with results from sports matches, but the goal stays the same: to build a predictive model and an investment (betting) strategy that beats our virtual bookmaker. We traditionally organize the hackathon together with the Intelligent Data Analysis team at FEL CTU, and the contest runs in two stages: an online qualification and a final round in Prague.

This year it was hockey. And you could tell!

After last year’s NBA, Czech hockey was up next. That is a completely different league, even for models. Hockey is among the hardest sports to predict: outcomes have higher variance, games are tighter, and when you add the fact that there is less available data than for overseas NBA, you get a task that is inherently trickier. That is exactly what this year confirmed.

By every measure, the assignment was the most challenging in the hackathon’s history so far. All the more impressive, then, that the winning teams managed to deliver solid virtual profits even in such a complicated environment. It turned out once again that the harder the problem, the better the ideas it draws out.

Simplicity wins the game

One interesting takeaway this year: despite the complexity of the data, the best-performing strategies were built on simplicity and elegance. The point was not to create the most complicated model on the planet, but to have a well-thought-out approach, clean signals from the data, and realistic handling of uncertainty.

Tonda Kříž, who competed last year as a student and now works with us as a Python developer, summed it up nicely: “Most teams went for simpler models, mainly logistic regression or decision trees. But we also saw more sophisticated feature engineering or Kalman filters. What was critical was working properly with risk: how much to bet based on how much money you have, how confident you are, what is allowed, and what might change during the competition. Those who stuck with fixed estimates often hit a wall.”

Tonda also warned about a classic trap: overfitting. A model that is “too smart” can start memorizing noise or random fluctuations in the training data, and then fails in real evaluation. Several teams burned out on that this year. By contrast, the teams that kept their models grounded and invested energy into meaningful features were the ones scoring points.

High-school teams among the best

Another highlight this year: two high-school teams made it into the finals. A team from Jan Kepler Grammar School even secured a fantastic 4th place. In a field packed with university teams, that is a result worth special mention. For us, it is a clear sign that this mix of mathematical and programming talent, plus healthy competitiveness, can thrive already at high school level, and that it makes sense to keep the hackathon open to them.

Thank you for taking this on with us

A huge thank-you goes to all participants who did not shy away from a hard assignment, and to the organizers and mentors who, under the lead of Gustav Šír, turned the weekend into an intensive lab of data, strategy, and well-aimed bets. Our collaboration with FEL CTU is crucial to us in the long run, and this year once again proved just how strong that connection is.

Congratulations to the winners!