A Forecast-Guided Hybrid Autoscaling Policy for Cost-Efficient, SLA-Aware Elastic Provisioning in Public Cloud Environments

Abstract

Reactive threshold-based autoscaling remains the default provisioning mechanism on most public cloud platforms, yet it reacts only after a service level objective is already at risk, while purely predictive approaches can under-react to sudden, unforecastable load spikes. This paper proposes a hybrid autoscaling policy that combines short-horizon linear-trend load forecasting with a low-latency reactive safety net that bypasses the scaling cooldown only when instantaneous utilization crosses an emergency threshold. The policy is evaluated in a discrete-event simulation against a fixed-step reactive baseline (modelled on common Amazon EC2 Auto Scaling step-scaling defaults) and a pure forecast-only policy, over a synthetic seven-day workload combining diurnal traffic, weekday/weekend modulation, and injected burst events. Under this workload, the proposed hybrid policy reduces total operating cost (infrastructure plus SLA-violation cost) by 4.7% relative to the reactive baseline while simultaneously reducing the SLA violation rate from 1.34% to 1.19% of time-steps, whereas the pure predictive policy increases both cost and violations. The results suggest that forecast-driven scale-out is only safely cost-effective when paired with a fast reactive fallback, and that the fallback's trigger sensitivity — not the forecasting model itself — is often the binding constraint on realized savings.