Informs Annual Meeting 2017

TD01

INFORMS Houston – 2017

Tuesday, 2:00 - 3:30PM

In this work, we develop methodology for the “one-shot” prediction environment, where each forecaster predicts a single outcome. Our aggregator arises from a mixture model of measurement error and information diversity, and is therefore, to the best of our knowledge, the first method that addresses both sources of variation simultaneously. Estimation relies on Bayesian statistics, leading to a posterior distribution of the consensus aggregate. Our methodology is illustrated on real-world and synthetic forecasting data. 2 - Quantifying Uncertainty using the Mean and Variance of a Collection of Point Estimates Asa Palley, Indiana University, 1275 E 10th St, Bloomington, IN, 47405, United States, apalley@gmail.com The wisdom of crowds—combining information from a collection of individual judgments—offers a useful method for quantifying an unknown variable. Averaging point estimates has proven to be effective in reducing error in the consensus estimate. Often, however, the decision maker requires an assessment of the full distribution. Using a Bayesian model of information, I show that the mean and variance of a variable of interest can be estimated as a function of the mean and variance of a collection of individual judgments, and derive the crowd consensus distribution. This provides a procedure that is easy to implement to estimate a predictive distribution from a collection of point estimates. 3 - A General Method for Addressing Forecasting Uncertainty in Inventory Models Dennis Prak, University of Groningen, Nettelbosje 2, Groningen, 9747 AE, Netherlands, d.r.j.prak@rug.nl, Ruud Teunter We derive a framework for addressing the estimation uncertainty, applicable to any inventory model and demand distribution. The (random) estimation errors are modeled and substituted into the inventory model, and the expectation of the objective function is taken with respect to the errors. When estimates are based on 10 observations, relative savings are typically between 10% and 30%. They are larger when estimates are based on fewer observations, when backorders are costlier, and when the lead time is larger. The method applies to any field where forecasts are used in optimization models. 4 - Estimation Error in Portfolio Optimization: Extra Constraints or More Diversification? Luis Chavez-Bedoya, Universidad Esan, Nestor Bermudez 191, Chorrillos, Lima, LIMA 9, Peru, l.chavezbedoya@northwestern.edu We study the relationship between extra constraints and diversification when reducing estimation risk in portfolio optimization. By adding a set of linear constraints to the original mean-variance portfolio problem, we prove that the portfolio with the minimum loss caused by parameter uncertainty consists on the combination of two sample portfolio with minimum risk and another one with zero exposure to a set of factors. Also, we claim that the best set of constraints corresponds to statements regarding the sample tangent portfolio. Finally, we give an explanation of the portfolio weights of Kan and Zhou (2007) and the portfolio rule of MacKinlay and Pástor (2000).

TD01

310A Decision Analysis in Public Policy Sponsored: Decision Analysis Sponsored Session Chair: Jonathan Welburn, RAND Corporation, 4570 Fifth Ave #600, Pittsburgh, PA, 15213, United States, jwelburn@rand.org 1 - Allocating Resources to Account for Multiple Levels of Uncertainty Lei Yao, Iowa State University, Ames, IA, 50011, United States, lyao@iastate.edu, Cameron MacKenzie Mathematical models to help public policy decision makers often have a great amount of uncertainty. This uncertainty includes uncertainty about parameters, models, decision maker risk attitudes, and even functional forms in the models. Policy makers may also be skeptical about solely relying on model recommendations with only optimal solutions. A solution to this uncertainty and decision makers’ skepticism is to find intervals as model outputs rather than a single optimal solution. This presentation will identify intervals for resource allocation models in which every objective function value evaluated by the interval differs from the minimum objective function value by a predetermined gap. 2 - Supporting Policy Decisions through Text Analytics Seifu John Chonde, Associate Operations Researcher, RAND Corporation, 1200 South Hayes Street, Arlington, VA, 22202, United States, schonde@rand.org, Timothy Stacey, William Marcellino, Ryan Brown RAND will discuss lessons learned in the development and application of multilingual text analysis tools to support policy decision makers. In recent work, we created a suite of data storage, natural language processing, and presentation tools. These tools are used across a range of tasks including lexical analysis, stance analysis, topic modeling, text classification, and advanced text clustering. The talk will focus on our latest developments in advanced text clustering. By walking through examples we will illustrate how text clustering is useful in public policy and highlight open research questions for the decision analysis community. 3 - Multinational Direct Investment and Political Uncertainty Omar Sherif Elwakil, Massachusetts Institute of Technology, 1 Amherst Rd., Cambridge, MA, 02139, United States, oelwakil@mit.edu Recent developments in the general equilibrium theory of multinationals emphasize the importance of multilateral considerations. Yet, existing explanations and corresponding estimations of FDI patterns have largely limited political and institutional investment impediments to a bilateral framework. Through the application of spatial econometric techniques, I show that the presence of both domestic and regional political uncertainty generate real options effects that lead to the delay or redirection of FDI. The magnitude and direction of these effects is conditional upon the host country regime type and the predominant multinational integration strategies in the region. 4 - Perceptions of Income Tax Evasion and Tax Moral: A Survey to Inform an Agent-based Model Raffaele Vardavas, RAND Corporation, Santa Monica, CA, United States, rvardava@rand.org We built an agent-based computational simulation model of income tax evasion. Within the simulation, individuals’ compliance behavior changes through an adaptation process based on their past experiences with audits and tax evasion penalties, their perception of the fairness in taxation rates and social interactions with people in their social networks. To inform the model we have conducted a survey on a nationally representative sample on the perceptions of tax fairness. We present how our model was informed by our survey and show some initial model results. 310B Probability Judgments Sponsored: Decision Analysis Sponsored Session Chair: Asa Palley, Indiana University, Bloomington, IN, 47405, United States, apalley@gmail.com 1 - One-Shot Forecast Aggregation under a Mixture of Measurement Error and Information Diversity Ville Satopaa, INSEAD, Paris, France, ville.satopaa@insead.edu TD02

TD03

310C Location Models for Emergency Service Applications

Invited: Tutorial Invited Session 1 - Location Models for Emergency Service Applications

Vladimir Marianov, Pontificia Universidad Catolica de Chile, Department of Electrical Engineering, Vicuna Mackenna 4860, Santiago, 782-0436, Chile, marianov@ing.puc.cl The term “emergency services” commonly refers to medical, firefighting or police services, although emergency repair can also be included in this category. All these services can be found in both public and private sectors. The design of police services requires mainly districting, dispatching, routing, and scheduling of police beats. Medical and firefighting services, on the other hand, mainly involve locating infrastructure and finding an initial or “idle” site where vehicles wait for an emergency to occur. Thus, the design of medical and firefighting systems benefit the most from the use of location models, and we concentrate in these two types of services. We do not perform a thorough review of the literature related to the application of location models to these services. Rather, we follow some of the successive model innovations that have addressed the main issues and added increasing degrees of reality to the modeling, with a didactic goal. We thus apologize to colleagues who have made significant contributions not included in this chapter.

374