Inspiration
At TAMU Datathon 2026, the House Edge challenge inspired us to ask a simple question: Can we uncover hidden patterns in seemingly random slot machine outcomes?
We wanted to go beyond guessing the next result and build a system that could estimate probabilities accurately, adapt to uncertainty, and improve through scientific experimentation.
What it does
All In is a machine-learning probability forecasting system that predicts the likelihood of five possible slot machine payouts: 0×, 1×, 2×, 5×, and 20×.
Instead of predicting a single outcome, our model generates a complete probability distribution using historical outcomes, lag-based patterns, and Bayesian inference.
Our strongest fully reproducible model achieved 53.53/60 points in local validation, substantially improving upon the competition's reference forecasts.
How we built it
We built All In using Python, NumPy, SciPy, GitHub, Claude AI, and ChatGPT.
We experimented with multiple statistical and machine-learning approaches before developing a specialized Bayesian forecasting pipeline.
One major breakthrough was separating the problem into two predictions:
What is the probability of receiving any payout?
If a payout occurs, what is the probability of each reward?
We also discovered hidden mathematical structure in the released payout probabilities, allowing us to simplify our model and improve accuracy.
Claude helped us implement, debug, and test models. ChatGPT helped us research published techniques, design experiments, and develop innovative approaches to overcoming performance plateaus. GitHub helped us track changes and maintain reproducible results.
Challenges we ran into
Our biggest challenge was overfitting. More complex models often appeared promising but performed worse under careful evaluation.
We also encountered numerical errors, limited historical data, changing machine behavior, and difficulties distinguishing genuine improvements from random fluctuations.
Breaking through accuracy plateaus required repeatedly analyzing our model's weaknesses, researching new approaches, and testing whether each change actually improved predictions.
Accomplishments that we're proud of
Achieved 53.53/60 on our reproducible local validation pipeline.
Developed custom Bayesian probability estimators and lag-based forecasting methods.
Discovered meaningful mathematical structure in the payout distributions.
Built automated validation and reproducibility checks.
Used AI-assisted research and development to rapidly explore advanced statistical techniques.
Most importantly, we learned to prioritize verified improvements over impressive-looking results.
What we learned
We learned that more complexity doesn't always mean better accuracy.
This project strengthened our understanding of Bayesian inference, probability distributions, statistical modeling, overfitting, and reproducible machine learning.
We also learned how to use AI tools effectively. ChatGPT and Claude accelerated our research and development, but independent testing and critical thinking were essential to building a reliable solution.
What's next for All In
Gamble with it and continue improving All In by exploring relationships between machines, detecting hidden changes in behavior, and developing more adaptive forecasting models.
We're especially interested in combining Bayesian inference with lightweight state-space models to improve predictions under uncertainty.
