Evaluating Amazon EC2 Spot Price Prediction Models Using Regression Error Characteristic Curve

Abstract

Amazon EC2 offers inactive virtual machines (VM) as spot instances at up to 90% discount. In return, the least expensive option requires the customers' usage to be tolerated with a low availability level agreement. Thus, many studies proposed forecasting and prediction mechanisms to asses in finding the best set of maximum prices. In this paper, we study the model's efficiency in predicting spot EC2 prices with focusing on assessing the performance of forecasting algorithms: RFR, XGBoost, k-NNR, and SVR. Model's evaluation is crucial for measuring the accuracy of predicted prices, thus, we select six metrics for evaluating the forecasting results. We used the top implemented metrics in the related work: MAPE, RMSE, MAE, and MSE. In addition, we assessed the spotted models using the Regression Error Characteristics (REC) curve and the Area under the curve (AUC-REC) in comparison to prior measures. Three aspects are considered while building the models: dataset time per year, training window as 1-day or 1-month ahead and instance location. The trained model applies the cross-validation technique to learn the ideal hyper-parameters that achieve the highest accuracy. However, except for the SVR model, our findings indicate it is unnecessary to use this technique to improve the algorithms' accuracy. Our results investigations display the REC curve and AUC-REC as a superior performance measurements for evaluating models over different accuracy-loss thresholds.