Overview

The Gilded Ace’s ten lowest-earning machines are not automatically its ten weakest games. A machine can produce low house win because of limited exposure, downtime, an unfavorable location, or ordinary payout volatility. Our dashboard separates these explanations before recommending replacement.

We built a reproducible Python, DuckDB, Parquet, and Streamlit pipeline covering January–June 2025. Machine performance is evaluated using theoretical win, coin-in per available hour, availability, and the difference between actual and theoretical win. Ratios are calculated from aggregated totals rather than averages of daily percentages. Machine scorecards classify poor results as primarily consistent with luck, downtime, location, payout behavior requiring inspection, or persistent weak-game demand. These labels are decision aids—not proof of mechanical failure or wrongdoing.

Time-series Model

We also estimated demand using a histogram gradient boosting model trained only on January–April data and tested on the untouched May–June period. The model used calendar and historical demand features available at prediction time and was compared with a constant training-period mean baseline. The model achieved a May–June MAE of 36.56 compared with the best baseline model MAE of 43.76. This validation informs the decisions given in the final recommendation while avoiding future-data leakage.

Downtime loss is reported as both gross machine-level theoretical win and a substitution-adjusted floor estimate. Because guests may move to nearby machines, we show sensitivity under multiple non-substitution assumptions rather than presenting one uncertain estimate as fact.

Outcomes

The final recommendations considered every machine as a potential candidate, along with consideration of each sub-category (zones, banks, etc.). After a final ten candidate machines were selected, a hypothesis test considered whether or not the set of action items could be financially appropriate, selecting only the tests with positive gains given pessimistic assumptions. In addition, addition action item hypothesis tests were considered on outliers, such as poorly performing banks. Again, only action items with positive gains were accepted here. Tested claims have a traceback sources backed with multiple tests, alternative options, results, and limitations. Actionable items are written to ./out/portfolio.csv

Our engine found these three most-important action items given our config assumptions:

  1. Add zone signage for Z-HTL (low confidence): This sub-section had the lowest floor coin-in area. The cost is $7,500 and the expected effect is $9,387: an expected profit of $1,887 over the next 30 days. How we'll know it worked: Zone-level coin-in per available hour vs other zones, plus entry counts where counter_status is valid.

  2. Inspect machine M-1158 (medium confidence): The machine scorecard indicates it is an outlier. The luck test Z-score is large (-4.11), which indicates a large deviation from expectation compared to its peers. The two-way effects test based on location showed it is not location that causes this outlier effect. The cost is $1,200 and the expected effect is $1,738: an expected profit of $538 over the next 30 days. How we'll know it worked: Verify meters and game setting against game_math; if a setting error is found, re-measure hold over 30 days and compare with its game-setting peers.

  3. Add zone signage for Z-B40 (low confidence): Second-lowest floor coin-in area. The cost is $7,500 and the expected effect is $7,891: an expected profit of $491 over the next 30 days. How we'll know it worked: Zone-level coin-in per available hour vs other zones, plus entry counts where counter_status is valid.

Other machines flagged or other action items did not have a net-gain over the next 30 days after total costs were considered. Changing model assumptions in config.py may change these outcomes.

Built With

  • dashboard
  • gradientboost
  • python
  • streamlit
Share this project:

Updates

Submission history