Author Archives: Nikos

Forecasting high-frequency time series with Neural Networks – an analysis of modelling challenges from increasing data frequency

S. F. Crone and N. Kourentzes, 2008, Proceedings of the International Conference on Data Mining, DMIN’08, Las Vegas, USA, CSREA, pp.37-42.

Prior research in forecasting economic time series with artificial Neural Networks (NN) has provided only inconsistent evidence on their predictive accuracy. In management science, NN were routinely evaluated on a set of well established benchmark time series of monthly, quarterly or annual frequency. In contrast, NN are accepted as a potent method for electrical load using time series that essentially display similar archetypical patterns of seasonality, trends, level shifts, outliers and calendar effects, but are of higher complexity and frequency. While this discrepancy has been attributed to the lack of a reliable methodology to determine the model parameters, recent research originating from econometrics and finance has indicated that high frequency data may pose different modelling problems compared to their low frequency counterpart and may hence benefit the use of dissimilar methods. This analysis aims to identify and contrast the challenges in modelling NN for low and high frequency data in order to develop a methodology tailored to the properties of the dataset. We conduct a set of experiments in three different frequency domains of daily, weekly and monthly empirical data of the same time series of cash machine withdrawals, using a consistent modelling procedure. The comparison against the naive model and exponential smoothing family models provides evidence that NN are suitable to predict high frequency data. Our analysis identifies a set of problems in modelling NN that arise in high frequency domain, mainly in specifying the input vector. To address these problems a different modelling approach is required between the low and high frequency data. Identifying these problems provides a starting point for the development of a unified methodology to forecast high frequency data with NN and facilitate revisions of the NN modelling approaches employed for low frequency data in management science.

Download paper.
Download presentation.

Related Posts

Automatic modelling of neural networks for time series prediction – in search of a uniform methodology across varying time frequencies

N. Kourentzes and S. F. Crone, 2008, Proceedings of the 2nd European Symposium on Time Series Prediction, ESTSP’08, Helsinki, Finland.

In time series prediction, modelling neural networks poses multiple challenges in specifying suitable input vectors, network architectures, and training parameters depending on the underlying structure of the time series data. The data properties are often determined by the frequency in which the time series is measured, such as low frequency data of yearly, quarterly or monthly observations, or high frequency data of weekly, daily, hourly or even shorter time intervals. As different time frequencies require distinct modelling heuristics, employing neural networks to predict a set of time series of unknown domain, which may exhibit different characteristics and time frequencies, remains particularly challenging and limits the development of fully automated forecasting methodologies for neural networks. We propose a methodology that unifies proven statistical modelling approaches based upon filters and best practices from previous forecasting competitions into one framework, providing automatic forecasting without manual intervention by inferring all information from the data itself to model a diverse set of time series of varying time frequency, like the ESTSP’08 dataset.

Download paper.
Download presentation.

Related Posts

Input variable selection for time series prediction with neural networks-an evaluation of visual, autocorrelation and spectral analysis for varying seasonality

S. F. Crone and N. Kourentzes, 2007, Proceedings of the 1st European Symposium on Time Series Prediction, ESTSP’07, Helsinki, Finland.

The identification and selection of adequate input variables and lag structures without domain knowledge represents one the core challenges in modeling neural networks for time series prediction. Although a number of linear methods have been established in statistics and engineering, they provide limited insights for nonlinear patterns and time series without equidistant observations and shifting seasonal patterns of varying length, leading to model misspecification. This paper describes a heuristic process and stepwise refinement of competing approaches for model identification for multilayer perceptrons in predicting the ESTSP’07 forecasting competition time series.

Download paper.

Related Posts