The 38th International Symposium on Forecasting (ISF) was held at the University of Colorado in Boulder. A large number of delegates from the Centre for Marketing Analytics and Forecasting attended and presented their latest research results.
John Boylan presented a work entitled “Multiple seasonal ETS models for multiple series” which exploits new methods for data when with the presence of double or triple seasonality. Such data are challenging to estimate, especially when data histories are short. In his work, John and his colleagues introduce new state space models with a single source of error and multiplicative components. They underline the procedures for forecasting multiple series with individual levels and trends, but multiple seasonal factors, with the flexibility to assume some of them are common across series. Optimal forecasting procedures and prediction intervals are derived, analytically where possible, and through simulations. Download John’s slides for more details.
Anna Sroginis co-organised an invited session on judgmental forecasting and adjustment. Anna’s talk elaborated on how algorithmic and qualitative information are interpreted when making judgmental forecast adjustments. To investigate these question, Anna conducts experiments that simulate a typical supply chain forecasting process that additionally provides qualitative information in presence of promotions. They found that participants tend to focus on several anchors: the last promotional uplift, current statistical forecast and contextual statements for the forecasting period. At the same time, participants ignore the past baseline promotional uplifts and domain knowledge about the past promotions. They also discount statistical models with incorporated promotional effects, hence showing lack of trust in algorithms. Download Anna’s slides for more details.
Also in the same session, Robert Fildes gave a talk on the dynamics of judgmental adjustments. There is little understanding on how experts adjust forecasts across time and whether their effectiveness is consistent across forecast horizons. Furthermore, it is unknown whether experts incorporate the same information, and in the same way, for different horizons. Robert and Nikolaos Kourentzes address these questions by investigating the dynamics of judgmental adjustments made by a UK manufacturer over a forecast horizon of 12 months. Their findings suggest that there is a clear shift in the type of information integrated into the statistical forecasts, particularly across forecast horizons where the magnitude, direction and the relative ‘value-added’ of the adjustments are horizon dependent. Download Robert’s slides for more details.
The talk of Nikolaos Kourentzes explores the topic of higher moments of information criteria for forecast selection and combination and proposes a modification in the use of the AIC and an associated procedure for selecting a single forecast or constructing combination weights that aim to go beyond the use of a simplistic summary statistic to characterise each forecast. He demonstrated that the approach does not require an arbitrary dichotomy between forecast selection, combination or pooling, and switches appropriately depending on the time series on hand and the pool of forecasts considered. The same procedure can also be applied to a wide variety of metrics, such as cross-validated errors as shown on a large number of real-time series from various sources. Download Nikos’ slides for more details.
Patrick Saoud gave a talk on the topic of information sharing in the presence of promotions in a supply chain. As many supply chains experience the Bullwhip effect, the effect of promotions on it is less researched. Patrick research examines the effect of promotions and other demand shocks on the performance of the different tiers of the supply chain. In particular, he studied the impact of promotions on forecasting accuracy, Bullwhip propagation and safety stocks for the participants of the supply chain. Furthermore, he also investigated the impact of different types of information sharing in this context on the Supply Chain and compared their performance regarding gains in forecasting accuracy. Download Patrick’s slides for more details.
Oliver Schaer presented research work on estimating the market potential pre-launch with search traffic. Typically, estimating the market potential of a new product is done by using expert judgment. However, there is substantial evidence that experts are biased and other ways to measure consumer interest such as surveys are expensive to gather. In contrast to this, pre-release buzz reflects the aggregate anticipation of consumers towards a new product and is readily available for example via Google Trends. Oliver proposes a method which is based on product generations that focus on the prediction of life-cycle sales. An empirical analysis on the case of video game sales shows that the method not only outperforms analogy based methods but also requires very few data points in order to be applicable. Download Oliver’s slides for more details.
In his talk, Ivan Svetunkov introduced a new framework for forecasting intermittent data with complex patterns. This framework is based on the intermittent state-space model developed by Ivan and John Boylan and the principles of the logistic regression. It allows not only capturing the complex dynamics in both demand sizes and demand occurrences (with potential trends, seasonality and special events) but also producing forecasts that incorporate these features. A large empirical evaluation on a dataset of a retailer shows the superiority of this new approach compared to existing intermittent demand forecasting methods such as Croston and TSB. Download Ivan’s slides for more details.
Dan Waller investigates the problem when retailers face variable data series when analysing sales of individual stock-keeping units. In his talk entitled “Hierarchical forecasting from a parameter estimation perspective”, he examines ways to improve estimation and forecasting by using hierarchical information. More specifically, he explores whether including combinations of hierarchies can lead to improvements not only in how the forecasts are adjusted but also in how the parameter estimates of those models might be simultaneously adjusted alongside them. The aim is a consolidated hierarchical forecasting approach which emphasises improved parameter estimation, but also reconciles forecasts in line with those adjustments. The evaluation examines the impact on both estimation accuracy and forecast accuracy. Download Dan’s slides for more details.