Nikos Kourentzes delivered the Keynote talk of the forecasting stream on the topic of “Uncertainty in predictive modelling”. The topic has substantial implications for estimation, model selection and eventually decision making. In that, Nikos first argued that model-based uncertainty should be accounted in prediction intervals and decision-making. This is because many forecasting techniques assume that the model itself is “true”, which almost always is not true. He then provided initial results from an approach to directly account for model selection uncertainty that leads to improvements in forecasting accuracy due to more robust model selection and combinations. Finally, Nikos pointed out that Multiple Temporal Aggregation, the topic he has been developing for the last five years, is one of the effective ways to address model uncertainty issue.

Patrick Saoud provided a presentation entitled “An empirical approximation for setting safety stocks”. The talk investigated the quantification of the uncertainty that stems from demand. In practice, the actual variance of forecast error is often underestimated, resulting in lower achieved service levels. Patrick highlighted shortfalls of the current standard method and introduced an intuitive approach to estimate the lead time demand variance. The superiority of the proposed method was demonstrated in an inventory simulation.

The two presentations are available for download here.