Find us on Facebook Find us on LinkedIn Follow us on Twitter Subscribe to our YouTube channel

Business Analytics Seminars

The seminars are on Fridays at 11am in Room 498, Merewether Building (cnr of City Road and Butlin Avenue), unless otherwise specified.

The seminar organiser is Laurent Pauwels.

Upcoming Seminars

7th Aug 2015 - 11:00 am

Venue: Seminar Room 498, Merewether Building H04

Speaker: Nathaniel Richmond, University of Iowa

Title: Decision-Dependent Uncertainties and Stochastic Incremental Network Design

In this talk, I will discuss two current projects. First, I discuss the stochastic incremental network design problem from an analytical standpoint. This problem comprises a series of optimization problems that one must solve over a planning horizon. In each time period, a network optimization problem (shortest path, max flow, etc.) is solved. Then a small change is made to the underlying network, i.e. an arc is added to the network. The goal is to minimize the sum of all costs in each time period by determining the order in which to build potential arcs. We consider a multistage stochastic program with uncertain arc construction times and discuss some specific analytical results on deterministic vs. stochastic solutions. The second project I discuss is a two-stage stochastic network resilience program with decision-dependent uncertainties. The original motivation for the problem is the protection of telecommunications or power networks against targeted attacks. Given an existing network, first the user decides which nodes to reinforce with a limited budget. Then a random event is realized (i.e. some nodes are destroyed), with the scenario probability distribution determined by the user's reinforcement decision. (This is the source of much of the problem's difficulty!) After the "attack," the user has a limited budget to repair damaged nodes. The objective is to minimize the expected value of the connectivity metric of the network after reinforcements, attack, and repairs. We discuss methods for managing the high computational complexity of decision-dependent uncertainties in this setting.

14th Aug 2015 - 11:00 am

Venue: SR498, Merewether Building H04

Speaker: Dr Nikolaos Kourentzes, Lancaster University

Title: Forecasting With Temporal Hierarchies

Using information from cross-sectional time series hierarchies often results in improved forecasting accuracy across all series. In this paper we introduce the concept of forecasting individual time series using temporal hierarchies. Temporal hierarchies comprise the observations of the highest frequency of a time series at the bottom level (say monthly), observations of the lowest frequency at the top level (annual) and in-between frequencies in the intermediate levels of the hierarchy (say, bimonthly, quarterly, semi-annual, etc.). By forecasting each of these components individually and then combining them using hierarchical time series approaches we observe two important results.

Firstly, we observe significant improvements in forecast accuracy at all levels of aggregation for the observed time series. Secondly, we generate reconciled forecasts across the different aggregation levels resulting to short, medium and long-run reconciled forecasts. We validate our results with an extensive empirical evaluation and perform a detailed simulation study in order to get an in-depth understanding of the forecast improvements we observe empirically.