Risk-Calibrated Hybrid Forecasting for FinOps-Driven Cloud Cost Optimization of AI Workloads

Abstract

This paper proposes a risk-calibrated forecasting framework that tackles the problem that existing cloud cost models are frequently concerned with providing point estimates and lack uncertainty awareness or operational decision support for FinOps-driven cloud cost optimization of AI workloads. The proposed approach is to build a cost signal at runtime from large-scale GPU cluster traces, partition the cost of the resources into the execution windows in which they are used, and create a next-hour cloud spend model based on temporal workload, GPU, CPU, memory, and rolling cost features. A hybrid ensemble forecaster is used in conjunction with conformal prediction for obtaining calibrated cost intervals and risk-aware budget-breach signals. Experimental results demonstrate good forecasting ability with the proposed hybrid ensemble model with MAE of 282.8970, RMSE of 360.1643, R2 of 0.8831, which are comparable to the Ridge Regression model and allow for uncertainty-based decision analysis. At 80%, 85% and 90% target coverage the conformal interval was successfully calibrated to PICP values of 0.8303, 0.8532 and 0.9128, respectively. The point forecast policy had the best F1-score of 0.6286 and the upper-bound policy recalled 0.9412 out of all the budget-breach events. The point forecast policy also had positive net-saving scenarios in 72.92 percent of the cases and a best saving of 3105.41 under the conformal upper-bound policy. The results highlight the relevance of the proposed approach, which combines runtime-aware cost modeling, conformal uncertainty calibration, and FinOps-oriented risk policies for reliable AI workload cost governance.