Abstract

As cloud and multi-cloud strategies become more common, managing costs, such as spot

pricing becomes a challenge. This thesis explores the possibility of using ARIMA-based and

neural network-based forecasting methods to predict spot prices in both Azure and AWS. One

finding is that the characteristics of Azure spot price history are unsuitable for these forecasting

methods. In AWS three different models were trained to predict on a test set. Among these, the

LSTM network was the top performer and the ARIMA model was the worst one. However, due

to the method used during the model training it is unclear whether the neural network models

perform well on unseen data.