Inspiration
Management at The Gilded Ace, a fictional 300-machine Las Vegas casino, planned to replace the ten lowest-earning slot machines. But earning little doesn't mean a machine is the problem: it can be unlucky, in a quiet spot, a low-payback game by design, newly installed, or often broken. We wanted to know whether the proposed list was right before any money was spent.
What it does
Our interactive dashboard tests each machine's low earnings against those explanations and answers all six of the challenge's questions with graphs, tables and plain-language summaries. It opens with a one-minute summary, then has pages for the machine review, visitor demand, downtime, loyalty members, loyalty rewards, a forecast, and a machine estimator. A $100,000 action plan includes uncertainty ranges and a way to measure success.
How we built it
We wrote the analysis in Python, querying the Parquet data with DuckDB and pandas, charting with Plotly, and presenting everything in Streamlit. A shared helper package keeps metric definitions identical across every page. Our key tests were:
- Luck: how far each machine's results fall from what its game math predicts.
- Location and exposure: comparisons against neighbouring machines, adjusted for time on the floor.
- Natural experiments: the Back Forty walkway closure and the May comp change.
- Forecasting: a time-split model whose protocol we committed to before writing any model code.
We also used Claude (Anthropic) through Claude Code to help plan and write code. Every number is checked by automated tests.
Challenges we ran into
- Raw earnings mix several causes, so we built an exact breakdown that adds up to each machine's gap.
- The data has traps: a 23-hour daylight-saving day, machines not yet on the floor, and carded play that covers only part of all play.
- Some results rest on thin evidence, such as only four machine moves, so we reported ranges and called unusual machines "inspection leads", not wrongdoing.
What we learned
A ranking can mislead when it ignores luck, location, and exposure. Our simple weekday-average forecast beat two fancier models, and being honest about uncertainty made our recommendations stronger.
Claims and Evidence Appendix
Listed as a separate tab in the database under Methods and Evidence.