On Forecasting Amazon EC2 Spot Prices Using Time-Series Decomposition with Hybrid Look-Backs

In this paper, we study forecasting through time series decomposition to predict Amazon Elastic Compute Cloud (EC2) Spot prices. To achieve this, we first decompose the Spot price history into time series components; each component, which can exhibit deterministic or non-deterministic qualities, is then separately forecast using different standard forecasting techniques and look back periods; and finally, the individual forecasts are aggregated to form the Spot price forecast. We compare our approach with several standard forecasting methods, namely Näıve, Seasonal Näıve, ARIMA, ETS, STL, and TBATS. From experimental results we make two observations: (i) none of the evaluated forecasting techniques yields consistently good results across all Spot markets, even though Seasonal Näıve can be considered the most robust when applied to markets with strong seasonal components, and (ii) our proposed technique performs comparably with, and in some cases, outperforms, the state-of-the-art forecasting techniques. The latter observation suggests that time series decompositionbased forecasting with hybrid look-backs warrants further investigation.