7th Econometric Institute PhD Conference

Start date

Friday, 11 Jan 2019, 12:00

End date

Friday, 11 Jan 2019, 17:00

Santander (M2-11)
Van der Goot Building
Campus Woudestein
Workshop: Machine Learning

On Friday, 11 January 2018, the Econometric Institute PhD Conference (EIPC) will once again take place. PhD candidates from all subfields in the Econometric Institute will present some of their current research, followed by panel- and plenary discussions.

The EIPC is an excellent occasion for staff members to see what problems the current PhD candidates are working on. In return PhD candidates get feedback on their research from experts in the field. In addition, the EIPC aims to stimulate interaction and knowledge-sharing between the various subfields within our institute.

Invited speakers

  • Nemanja Milovanovic
  • Jochem Oorschot
  • Weina Ma
  • Anoek Castelein
  • Lisanne van Rijn
  • Sebastiaan Vermeulen


13:20-13:30Opening EIPC
13:30-14:00Nemanja Milovanovic
14:00-14:30Jochem Oorschot
14:30-14:45Coffee break
14:45-15:15Weina Ma
15:15-15:45Anoek Castelein
15:45-16:00Coffee break
16:00-16:30Lisanne van Rijn
16:30-17:00Sebastiaan Vermeulen
17:00 onwardsDrinks. Celebrating the start of the new year in the Paviljoen


Please fill in the registration form to indicate your presence at the lunch, the conference and a drink. You can also specify dietary preferences and restrictions on the form.

If you wish to change or cancel your registration, please contact the organisation via eb-secr@ese.eur.nl


    • Motivated by an aerospace service provider, we consider a make-to-stock company’s stock rationing and order decisions. We take into account the real time information on replenishment orders and number of waiting customers from different classes. The company faces two customer classes: (1) contract customers whose demands are practically always accepted and fulfilled, and (2) spot customers who arrive over time and are rejected, satisfied, or put on the waiting list. The spot customers put on the list may also leave if they have waited too long. We model the problem as an infinite horizon Markov decision process. The supply lead time has an Erlang distribution. The state space is comprised of the number of spot customers on the waiting list, the changing status of the replenishment order, and the inventory level. The last two are aggregated by a state variable called work storage level. The objective is to minimize the total discounted cost. Our numerical study shows the optimal rationing policy is a sequence of monotone critical work storage levels dependent on the number of waiting spot customers. We also find that the optimal policy outperforms policies which ignore the real time information on replenishment orders or the option of putting spot customer on the waiting list.

    • Over the past decades, containerships have rapidly grown in size. Nowadays, ships are being built that can accommodate over 22,000 TEU. This trend of continuously ordering bigger and bigger ships in order to pursue economies of scale has led to a market overcapacity. As a consequence, liner shipping companies have difficulties to capture enough demand to make these bigger ships cost-effective. Combined with a highly competitive market, liner shippers in general only see small profit margins, if any. In our research, we investigate the seemingly irrational behavior of continuously ordering bigger and bigger ships. We model the problem as a Stackelberg competition, where leader and follower decide on their rates and their liner network. We look at two settings. In the first setting, leader and follower are identical liner shipping companies. In the second setting, we put an upper limit on the follower’s fleet, such that they can only use ships with a capacity lower than a threshold. First results indicate that by restricting the follower’s fleet, the leader gains advantage due to being able to reap economies of scale. This advantage disappears, however, in the unrestricted case, and can even turn into a disadvantage.

    • Decentralized decision-making processes are often preferred in development organizations, but the corresponding lack of coordination may lead to sub-optimal outcomes. The extent of this has not yet been quantified. We investigate (de)centralized time allocation policies for family planning outreach teams in Uganda and show that decentralized policies perform near optimal.

    • In this paper, we develop a model to describe the choices of respondents during discrete choice experiments. The aim of the model is to infer the preferences of respondents while accounting for both heterogeneity across respondents in these preferences as well as in potential learning, fatigue, and boredom effects. Learning, fatigue, and boredom effects refer to a dynamic process that a respondent may go through while filling in a survey. When learning occurs, responses will become closer to a respondent’s ‘true’ preferences as the survey progresses. When fatigue or boredom occurs, responses will increasingly deviate from the ‘true’ preferences. 

      We develop a hidden Markov multinomial logit model that captures heterogeneity in learning, fatigue and boredom effects through time-varying error variances that follow first-order Markov switching processes. Intuitively this means that we assume that during the choice process, each respondent goes through a number of phases, where each phase is marked by a different error variance. For example, there is a learning phase with a small error variance and a bored/fatigued phase with a large error variance. We account for preference heterogeneity using a discrete finite mixture approach. For model inference, we rely on a Bayesian Markov chain Monte Carlo (MCMC) approach.

    • The decision to rebalance is a confounding factor for both the risk and return of the portfolio. This implies that in portfolio selection problems, the decision to rebalance should play a role equivalent to determining the portfolio weights. To tackle this issue, we extend Markowitz' mean-variance analysis to a finite-time multi-period setting, and we bridge the gap with growth-optimal portfolio selection.

      Our online learning approach is robust against model misspecification and allows for arbitrary temporal dependence, such as hypes and nonlinear causal relationships.

    • A common method for estimating the extreme value index in Extreme Value Theory is to divide a sample into k blocks and fit the sample of k block maxima to the Generalized Extreme Value distribution --- the block maxima (BM) method. The BM estimator generally changes under permutations of the observations because it changes the sample of block maxima. We propose the All Block Maxima (ABM) estimator in the case of i.i.d observations. The ABM estimator essentially weights the order statistics by the probability that it is a block maximum under random assignment of the observations to any block, and is itself permutation invariant. We study the ABM’s asymptotic properties and whether the estimator is more generally applicable than either the BM method or the common Peaks Over Threshold method

More information

For more information please contact Sebastiaan Vermeulen or Rolf van Lieshout, eipc@ese.eur.nl