Inspiration
Management at The Gilded Ace, a fictional 300-machine casino, planned to replace its 10 lowest-earning slot machines. But a machine can earn little because it is unlucky, in a quiet spot, or often out of service. We wanted to know if replacing them would help.
What it does
An interactive dashboard that follows the money from the whole floor down to each machine, then prices what to do in July. You can change the dates, the view (day to quarter) and the zone.
Finding: only one of the ten, M-1158, has a machine-specific problem. It pays on 38.5% of spins against 31.5% published. The other nine sit in a quiet zone, run low-margin games, or were not on the floor all period. Luck cost the ten about $63k.
Plan ($38,700 of $100,000): inspect M-1158 ($1,200), run two promotion weekends ($30,000), and put up signs in Back Forty ($7,500). Expected net gain is about $65k (range −$1.7k to +$151k). We keep $61,300.
How we built it
Python (pandas, DuckDB, scikit-learn) prepares the data, and a single HTML page with Chart.js shows it. For each machine we split its gap behind a typical machine into days on the floor, downtime, spot and game, the machine itself, and luck, using expected win $= \text{wagered} \times (1 - \text{RTP})$.
Two models, trained on January–April and tested on May–June:
- A floor forecast (weekday pattern × recent level): 4.9% error against 16.4% for the average day.
- A machine-level model (gradient-boosted trees) that estimates what a slot would earn with a different game.
Challenges
- Separating luck from location and tenure, because raw totals mislead.
- Reporting honestly: a simple zone × game × weekday table was about as accurate as our machine model, so we use the model for what-if questions, not as a claim of superiority.
- Keeping the dashboard simple for a manager.
- Leaving out source tables whose redistribution terms are not cleared.
What we learned
A low rank is not the same as a bad machine: luck, location and time on the floor explained nine of ten. We also learned to say what we could not prove. The effect of a different game, and of promotions, rests on few observations.
Links: dashboard [https://ghodkeadityarao-tamu.github.io/gilded-ace-dashboard] · code [https://github.com/ghodkeadityarao-tamu/gilded-ace-dashboard]. The AI-use disclosure and Claims & Evidence appendix are attached.
Built With
- chart.js
- css3
- duckdb
- github
- html5
- javascript
- numpy
- pandas
- pyarrow
- python
- scikit-learn
- statsmodels
