The conference was held in Santander, Spain and is organised by the International Institute of Forecasters (IIF). As a featured speaker, Robert Fildes, gave a presentation on ‘Research in Practice’. His theme was the importance of understanding the role of organisational practice for stimulating the development of new theories and models in forecasting. Robert showed how research in the field of forecasting has evolved over the years, pointing out that studies have often limited practical impact. There are under-researched fields such as structured judgment or grounded applications that use real-time data and include organisational context. The talk then illustrated what researchers can learn from practice with an example of analysing and improving an S&OP process. Concluding in suggestions how to bridge the gap between research and practice.

John Boylan presented the research conducted together with Zied Babai, Mona Mohammadipour and Aris Syntetos. His talk was on ‘Reproduction of Forecasting Methods in the M Competitions’. The authors reproduced the results of M1 and M3 using simple methods from the original study. They found some differences in forecasting at longer horizons. John concluded that the transparency of methods has not developed as fast as it should have since M competitions were published. Given this findings greater clarity on several aspects of forecasting such as error measures and seasonal adjustment procedures is needed.

Another featured talk was given by Nikolaos Kourentzes on ‘Forecasting with Temporal Hierarchies’. Nikos has also delivered a second presentation on ‘Measuring forecasting performance: A complex task!’ (joint research together with Ivan Svetunkov and Juan Ramon Trapero). The first presentation was given to the practitioner track of the conference and aimed to introduce the basic idea of temporal hierarchies for forecasting and highlight the forecasting and business challenges they can help to address. The second research presentation introduced a new error metric to evaluate forecast performance. Although there has been substantial research on accuracy metrics, there has been very limited work on bias metrics that has an equally important dimension of performance. The proposed metrics provides a symmetric and robust, scale independent bias measure and also contains the information about accuracy. An interesting aspect of the proposed metric is the informative visualisation of both components.

Ivan Svetunkov presented his research on ‘Model Parameter Estimation with Trace Forecast Likelihood’ (work in collaboration with Nikolaos Kourentzes). He showed that when multiple steps ahead estimators are used on univariate models, they impose shrinkage on parameters, sometimes forcing models to become deterministic. Ivan then proposed an alternative approach that leads to more accurate estimates and does not lead to overshrinkage of parameters.

Oliver Schaer presented on the topic of ‘Forecasting demand with internet search data and social media shares’ (supervised by Nikolaos Kourentzes and Robert Fildes). Several studies found information from online sources to be valuable leading indicators but few have demonstrated its usefulness in supply chain forecasting. Results from an initial experiment with physical online video games sales and search traffic, however, suggest no additional benefits. Questioning whether there is sufficient lead time present between search query and purchase, Oliver suggested that search traffic information might be still valuable on different product categories or within the area of pre-launch forecasting.

Daniel Waller made a presentation on ‘Modelling multiple seasonalities across hierarchical aggregation levels’. He discussed the application of high frequency forecasting models to retail company data and showed that no one has applied these models to such a data before. He then demonstrated on a time series example that such an approach holds promise in retail, pointing out that univariate models seem to perform better in short term, while multivariate ones may bring benefits to forecasting in long term.

Matt Weller gave a talk on ‘Temporal aggregation and model selection: an empirical evaluation with promotional indicators’ (research together with Sven Crone and Robert Fildes). His research compares bottom-up (weeks to months) and top-down (months to weeks) approaches taking promotions into account. The results of Matt's research show that there are substantial gains in forecast accuracy when using a bottom-up forecasting with explanatory variables. This implies that manufacturers should examine weekly forecasting when promotions with high impact and intensity is present.

Download the slides:

• Research in Practice - R. Fildes
• Reproduction of Forecasting Methods in the M Competitions - J. Boylan
• Forecasting with Temporal Hierarchies - N. Kourentzes
• Measuring forecasting performance: A complex task - N. Kourentzes
• Model Parameter Estimation with Trace Forecast Likelihood - I. Svetunkov
• Forecasting demand with internet search data and social media shares - O. Schaer
• Modelling multiple seasonalities across hierarchical aggregation levels - D. Waller
• Temporal aggregation and model selection: an empirical evaluation with promotional indicators M. Weller

Furthermore, the following ISF presentations from visiting LCF researchers are available for download:

• Developing intelligent tutoring support for forecasting‌ - D. Barrow
• Exploring the sources of uncertainty: why does bagging work?‌ - F. Petropoulos
• Variable Selection for long-term forecasting using temporal aggregation‌ - Y.R. Sagaert
• Optimal combination of volatility forecasts to enhance solar irradiation prediction intervals estimation‌ - J.R. Trapero