Presentations from the ISF 2017 conference
30 August 2017
A large delegation from the Centre for Marketing Analytics and Forecasting presented their latest research at the premier event for business forecasting in Australia.
The 37th International Symposium on Forecasting (ISF) was held in Cairns between 25th to 28th of June. The centre presented on a broad spectrum of forecasting-related topics.
Robert Fildes chaired the practitioner track on forecasting in retailing. In that session he first provided a short overview on retail forecasting, summarising research issues in the area. In particular recent research of dealing with the many variables relevant for SKU level forecasts and speculates about the effects of the growth of internet sales and information from social media.
Robert also presented findings from a collaboration work with Patricia Ramos who is visiting researcher of CMAF in Lancaster. The presentation entitled “Characterizing retail demand with promotional effects for model selection” provided evidence on the accuracy of various forecasting methods, both univariate and multivariate and gives preliminary results on the benefits and costs of selection.
A further session on hierarchical forecasting and applications was chaired by Nikos Kourentzes. He presented a joint work with Rickard Sandberg and discussed the implicit connections enforced by hierarchical time series forecasting, between the nodes of the hierarchy, contrasting them to VAR models that capture these connections explicitly.
Ivan Svetunkov, in collaboration with John Boylan, presented a new approach for forecasting intermittent and non-intermittent data. This approach allows capturing trends in both types of data. In an empirical experiment, the method was applied to a data with different patterns and outperformed all competing forecasting models.
Anna Sroginis, asked the question “Can forecasters can interpret salient information when making judgmental adjustments” at the session of judgmental adjustments. Her research shed light on the experimental design that allows investigating how forecasters interpret additional information when making judgmental adjustments, i.e. how experts perceive the usefulness of information and implement it in their adjustments when such soft information is provided.
In the presentation on using explanatory variables from social networks and search traffic to forecast demand, Oliver Schaer raised some concerns about their usefulness. The results from an empirical experiment, are in a stark contrast with the literature and found that established univariate forecasting benchmarks, such as exponential smoothing, consistently perform better than more sophisticated statistical models.
Daniel Waller presented his latest research, motivated by the task of sales forecasting at the most disaggregate level, an important topic for inventory management purposes. In his presentation, Daniel compared approaches which inherit information from more aggregate levels of the sales hierarchy, and, through simulation, evaluated the impact of hierarchies on both parameter estimation and forecast accuracy through simulation.
Finally, Yves Sagaert extended his research on macroeconomic leading indicators for tactical sales forecasting. In his presentation, he reviewed the prediction intervals of forecasts that make use of leading indicators. Yves investigated the effects of reconciliation across the product hierarchy. He showed that traditional weighting in forecast reconciliation might not be optimal for forecasts formulated with leading indicators, and argued for different weighting schemes.