Hardware failure prediction in electronic gaming machines using time-series data

Abstract

Modern hardware systems produce continuous telemetry that can reveal early signs of performance degradation and emerging failures. This thesis explores whether forecasting hardware telemetry variables, such as CPU utilization, memory usage, etc., can support proactive maintenance. Using 408 GB of Prometheus data, collected from three electronic gaming machines, we curated a data dictionary, ranked these variables, and selected the top five for detailed forecasting. Three time series forecasting models (ARIMA, Prophet, and Chronos) were evaluated under both univariate and multivariate settings. Performance was assessed using Mean Absolute Percentage Error (MAPE). Results show that univariate accuracy varies by metric; no single model dominates across all metrics. When contextual features are added, forecasting accuracy improves; Prophet achieves the lowest error. These findings demonstrate that telemetry can be reliably forecasted and provide strong baselines for early detection of abnormal hardware behavior, reducing downtime and extending hardware lifespan.