Casino Manager
Dashboard: https://thajiyev.github.io/dashboard-royale/ · Code: https://github.com/THajiyev/dashboard-royale
The bottom 10
Casino Manager found that the 10 lowest earners shouldn't be replaced. Most of the performance is explained by the poor location in Back Forty. M-1158 was the only machine flagged for inspection, but it was not confirmed as faulty.
Demand
Demand follows the day of the week and events such as sports, festivals, conventions and new banks. The walkway closure and a visitor-driven June slowdown also cut demand.
Downtime
Downtime cost around 5021 hours and $102K in theo (expected house win). Downtime happened every month, but it was the worst in May.
Players
Tracked players wagered $163.8M and got back $148.2M (a net loss of $15.6M). The top 10% of losers are responsible for 53% of losses.
Comps
Comp cost was ≈$191K in total. On May 1st, points per comp halved without providing extra play, and raised cost by 9%.
Model
For the ML section, I used a Poisson GLM to predict each machine's play per hour. I trained on January to April and tested on May to June. I improved the MAE by 28% compared to the baseline of zone averages ($53.1 vs $38.2). This model helps decide if moving a machine is worth it.
July plan
When planning out the month of July, I would spend $48.3K out of the $100K and keep $51.7K.
- Promotions ($30K). I would spend $30K on 2 promotion weekends. Past promotions added $8.4K and $26.1K of theo a day. The uncertainty comes from the small sample size of two events, but the outcome is easily measured by comparing to the July forecast (see appendix).
- Machine moves ($10.8K). I would move 6 machines for $10.8K. Both our ML model and past record of a move adding $342 a day support this idea. Uncertainty comes from our ML model's coefficient range. The outcome is easily measured by comparing the moved and unmoved machines' usage rates.
- Back Forty signs ($7.5K). Lastly, I would purchase new signs for Back Forty for $7.5K. Closing the walkway to the area cut its play by around 31%, so improving access should help. It is uncertain since this is mostly an assumption based on a logical connection. The outcome is measured by comparing this area to others.
Appendix
AI Statement
I utilized GitHub Copilot and Claude Code for planning, coding, and editing my documentation. I used ChatGPT to generate the website logo. Claude proofread and edited my writing. Claude helped me write the Claims & Evidence table from the findings already documented on the dashboard's Evidence page, including the alternative explanations and limitations, and I checked each row.
Another ML model (Optional Read)
I also used time-series to forecast theo for the day. My baseline was the average for each day of the week using data from January to April. The model used weekday averages and lifts for different event types. Tested on May to June, it cut the MAE by 36% ($6699 to $4290). This gave the July forecast that the plan's results are measured against.
Claims & Evidence
Every table is in data/, and scripts and model outputs are in the repository. The dashboard's Evidence page recomputes the headline claims from raw data (python scripts/reproduce_claims.py).
| # | Claim | Table / query | Result | Alternative explanation | Limitation |
|---|---|---|---|---|---|
| 1 | Most of the bottom 10 are low because of location | machine_day: SUM(house_win_cents) per machine on active days, lowest 10, zone from prepared_daily_view.zone_name (zone on most days) |
9 of the 10 sit in Back Forty | Looked for other similarities among the machines. | N/A |
| 2 | M-1158 is the only machine flagged for inspection | machine_day (house_win, coin_in, theo, paid_spins, payout_positive_spins) + game_math.hit_probability: observed vs configured hold, win-rate z per machine, Benjamini–Hochberg across 310 machines |
Hold 3.0% vs 10% configured, win-rate z = +124, only machine flagged after correction | Compared the expected and actual hold rate to see if the difference is statistically significant. | Statistical test, so results might not be 100% correct |
| 3 | Demand follows the day of the week and events | property_day.theo_usd + events.event_type, Jan–Apr: event-day theo ÷ the normal level for that weekday, averaged by event type |
Sports and festivals +17–20%, conventions +3% | Compared days with events to those without. | Small sample size (2 sports events, 3 conventions) |
| 4 | New banks raised demand | outputs/model_change_impacts.csv, bank B-51 rows (analysis/impact.py): new machines' theo minus cannibalisation of existing Pit machines |
+$5.3K theo a day (90% CI $4.6K–$6.0K) | Play could have just moved from existing Pit machines, so that cannibalisation is subtracted. | One new bank |
| 5 | The walkway closure cut demand in Back Forty | prepared_daily_view (coin_in_usd, available_hours, zone_id): daily log(coin-in per available hour), Back Forty minus 5 control zones, OLS on closed + reopened + weekday |
−30.9% while closed (Mar 3 – Apr 6), recovered after, whole floor +1.6% (90% CI −0.2% to +3.4%) | The whole floor could have slowed in March, but placebo zones moved only −2% to +3%. | One closure, one zone |
| 6 | A visitor-driven June slowdown cut demand | property_day.theo_usd, events (event days dropped): Σ weekday-baseline forecast ÷ Σ actual − 1 on ordinary June days, locals vs visitors split on the Demand page |
June 6.5% below weekday norms, mostly visitors | Heat could explain it, but 40°C+ days account for about 1 point at most. | Visitor split covers carded play only |
| 7 | Downtime cost about 5021 hours | machine_day: Σ (scheduled_seconds − available_seconds) over active machine-days |
5,021 hours, 0.4% of scheduled time, 98% fault repairs | Downtime could explain the bottom 10, but Back Forty had the fewest fault hours. | Meter-based, up-but-unplayed machines not counted |
| 8 | Downtime cost $102K in theo | outputs/phase0_downtime_lost_play.csv (analysis/run_phase0.py): down hours × the machine's theo per hour |
≈ $102K theo, Jan–Jun | Down machines' play could move to neighbours, so the true loss may be lower. | Assumes a down machine would have earned its usual theo per hour |
| 9 | Downtime was worst in May | prepared_daily_view.downtime_hours by month |
May 1,017 h (highest), June 631 h (lowest) | May has one more day than June, but that is far smaller than the gap. | Six months, no repeat season |
| 10 | Tracked players wagered $163.8M, got back $148.2M | sessions (carded play), Jan–Jun: Σ coin-in, Σ payout |
$163.8M wagered, $148.2M paid back, net loss $15.6M | Gaps against machine totals reflect uncarded play, not broken data. | Carded play only (about two thirds of coin-in) |
| 11 | The top 10% of losers account for 53% of losses | player_period: tracked_payout_usd − tracked_coin_in_usd per player, losers sorted by loss, share held by the top 10% |
53% | Bad luck could explain it, but expected losses are even more concentrated. | Carded play only, describes outcomes, not targets |
| 12 | Comps cost about $191K | player_period.comp_cost_usd, Σ Jan–Jun |
≈ $191K ($30.8K a month Jan–Apr, $33.7K May–Jun) | Menu prices would overstate it about fourfold, so actual cost is used. | Carded guests only |
| 13 | On May 1 points per comp halved | loyalty_ledger redeem events: median points per $1 of comp value, before vs after May 1 |
−53% (15 → 7 points per $1), every tier | The menu could have changed, but items kept the same face values. | Every member at once, no control group |
| 14 | Comp cost rose 9% | player_period.comp_cost_usd: May–Jun monthly ÷ Jan–Apr monthly − 1 |
+9.4% (≈ $2,900 a month) | Guests could have played more, but theo fell while comp cost rose. | Only two months after the change |
| 15 | The cut brought no extra play | player_period (theo_usd, home_region, comps_redeemed): theo May–Jun ÷ Mar–Apr − 1 |
Locals who used comps −2.5%, all locals −1.2% | A seasonal dip could hide a lift, so comp users are compared with all locals. | Locals only |
| 16 | The Poisson GLM beats zone averages by 28% | prepared_daily_view coin-in per available hour (analysis/machine_effects.py): GLM on bank + theme + weekday + events + closure, trained Jan–Apr, tested May–Jun |
MAE $38.2 vs $53.1 | A lucky test window could flatter it, but it is scored on unseen May–Jun data. | One two-month test window |
| 17 | The time-series forecast cuts error by 36% | property_day.theo_usd: weekday-mean baseline vs event-type lift model frozen on Apr 30, scored May 1 – Jun 30 (analysis/experiments/iter5/holdout_daily.csv) |
MAE $6,699 → $4,290 | Model shopping could explain it, but the model was picked before May was opened. | One 61-day window, 90% bands too narrow (caught 75%) |
| 18 | The plan costs $48.3K | action_catalog (unit_cost_cents, recurring_cost_cents_per_day, lead_days), GUIDE budget formula |
$48,300 spent, $51,700 kept | Replacing the bottom 10 instead, but the one replacement in the data added only $4 a day. | Catalogue costs and lead times taken as given |
| 19 | Past promotions added $8.4K and $26.1K of theo a day | outputs/model_change_scorecard.csv (analysis/impact.py): floor theo on promotion days vs expected level |
Locals Week +$8.4K a day (90% CI $3.4K–$13.3K), Winter Slot Tournament +$26.1K ($17.2K–$34.6K) | Busy weekends could lift anyway, but the estimate controls for weekday. | Two promotions only |
| 20 | A past move added $342 a day | outputs/model_change_impacts.csv, "Move M-1152 … NET" row: direct effect + spillover on bank-mates |
+$342 a day (90% CI $188–$497) | The move could have just pulled play from bank-mates, so spillover is subtracted. | One past move |
