Technical Details: The Statistical Algorithms Behind Our Dashboard
One important distinction: Gemini doesn't perform our statistical calculations, and Tiger Cloud isn't just storing data. We built the analytical methods separately, use Tiger Cloud to query historical hourly data efficiently, and let Gemini retrieve and explain the results.
Here are the main algorithms we implemented:
Leave-One-Out Matched Peer Benchmarking: We compare each machine against other machines with the same game configuration, location zone, month, and weekday. The target machine is excluded from its own benchmark, and performance is normalized by available operating hours. This helps separate low demand from limited exposure.
Payout Anomaly Detection (Z-Test + Benjamini–Hochberg): We calculate expected house win using each game's configured return-to-player rate. We estimate payout variance from squared bet amounts and game volatility, calculate a standardized Z-score, and apply Benjamini–Hochberg false-discovery-rate correction across machines. This flagged M-1158 with Z = −4.57 and adjusted q ≈ 0.00151, without claiming a malfunction.
Poisson Regression for Machine Reliability: We model fault counts using Poisson regression with active machine-days as an exposure offset. We control for manufacturer, game configuration, zone, month, and weekday, using quasi-Poisson covariance to account for dispersion. Ridgeline's adjusted fault rate was approximately 3.56× Halcyon's.
Fixed-Effects Regression for Location Changes: We analyze wagering per available hour using weighted log-rate regression with date and zone-by-configuration fixed effects. We use Newey–West standard errors with a seven-day lag to account for temporal correlation. This supports the Back Forty demand finding while acknowledging that the relationship isn't necessarily causal.
Block Bootstrap Confidence Intervals: We resample observations in seven-day circular blocks within months to estimate uncertainty in machine performance and contribution ratios. This preserves some weekly dependence that ordinary independent resampling would ignore.
Chronological Forecast Validation: We select a weekday-based forecasting model using earlier training and validation periods, then freeze predictions for May–June. Its daily-wager MAE was $68,699 versus $218,258 for a constant-mean baseline, a 68.5% reduction. We also evaluated additional calendar-based models retrospectively.
Loyalty Mix Standardization: We match loyalty redemptions to POS transactions and calculate actual economic costs. We then hold January's item-and-tier mix fixed to distinguish changes in point pricing from changes in what customers redeem.
Constrained Portfolio Enumeration: We generated 20 evidence-based action bundles and exhaustively evaluated 968 feasible portfolios under the $100,000 budget, three-action-type limit, lead times, costs, and incompatibilities. This produced our $0 financial recommendation and a separate $13,900 experimental strategy under a different objective.
Tiger Cloud Time-Series Aggregation: We imported 1,321,908 machine-hour records into a TimescaleDB hypertable with seven-day time partitions and hourly continuous aggregates. Utilization is calculated as
SUM(occupied_seconds) / SUM(available_seconds), rather than averaging percentages. We also rank one-to-four-hour maintenance windows by estimated theoretical contribution at risk. In measured Tiger Cloud tests, continuous aggregates made our full-property SQL query approximately 7.1× faster than scanning raw hourly data.
Gemini connects these findings through constrained function calling and verified evidence retrieval. It chooses which calculations and results are relevant to a question, while the underlying algorithms remain responsible for the numbers.
The goal was to make complex statistical analysis both reproducible and understandable—not to have an LLM guess what the casino should do.