Amazon EC2 Spot Price Prediction Using Temporal Convolution Network

Abstract:
Cloud service providers sell some idle resources at dynamic prices. These resources are called spot instances. The price of Spot instances is affected by the relationship between market supply and demand. Generally speaking, the price of spot instances is much lower than that of on-demand instances or reserved instances. The prediction problem of spot instances is very important for users. Being able to predict the price with high precision can enable users optimize their bid of spot instances so that they can run their tasks more economically. Cloud service providers will not announce their specific pricing mechanism, but AWS will give the historical price of spot instances in the past 90 days. By transforming the price prediction problem into a sequence data analysis problem, we can use the TCN(temporal convolution network) to predict the future price. The experimental results show that the TCN prediction model outperforms the traditional machine learning model and other models such as CNN and LSTM in MAE, MAPE and RMSE.