No abstract

No abstract

No abstract

Abstract. Most approaches to forecasting time series data employ onestep-ahead prediction approaches. However, recently there has been focus on multi-step-ahead prediction approaches. These approaches demonstrate enhanced prediction capabilities. However, multi-step-ahead prediction increases the complexity of the prediction process in comparison to one-step-ahead approaches. Typically, studies in the examination of multi-step ahead methods have addressed issues such as the increased complexity, inaccuracy, uncertainty, and error variance on the prediction horizon, and have been deployed in various domains such as finance, economics, agriculture and hydrology. When determining which algorithm to use in a time series analyses, the approach is to analyze the series for numerous characteristics and features, such as heteroscedasticity, auto-correlation, seasonality and stationarity. In this work, a comparative analysis of 20 different time series datasets is presented and a demonstration of the complexity in deciding which approach to use is given. The study investigates some of the main prediction approaches such as ARIMA (Autoregressive integrated moving average), NN (Neural Network), RNN (Recurrent neural network) and SVR (Support vector regression), which focus on the recursive prediction strategy and compare them to a new approach known as MRFA (Multi-Resolution Forecast Aggregation).

Highlights This paper presents a new method for generating 50K diverse synthetic time series. We present a discussion on time series characteristics and metrics with a view to understanding time series diversity. We developed a robust framework for validating diversity in synthetic time series generation.