Dynamic Spot Price Forecasting Using Stacked LSTM Networks

Cloud computing has gained more attention and emerged as a significant model for provisioning resources to the business industries. Maximizing profit and revenue from unused resources are the most trending goals of any cloud service provider. Amazon web services offer spot instances at very substantial down prices. The instances are offered through a bidding mechanism. One of the challenging tasks is to predict the spot price because spot users always want to make cost-effective decisions. Amazon provides the past 90 days of pricing history to assist its users in making bidding decisions. Accurate prediction reduces the risk of overbidding and out-of bid failure situations. In this paper, a 2-layer stacked LSTM model is used as a learning model for prediction purpose. We explore here the use of stacking the LSTM for time series prediction. The model is implemented using Keras and tested using the Amazon spot pricing dataset. Our results show that our stacked LSTM model provides better accuracy for next day spot price prediction.