Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MB39

strategizing about reports. Proper scoring rules incentivize truthful reporting if all forecasters are paid according to their scores. However, incentives become distorted if only the best-scoring forecaster wins a prize. We introduce a truthful forecaster selection mechanism, and lower-bound the probability that our mechanism selects the most accurate forecaster. 2 - The Effectiveness of Trimmed Prediction Polls in Time Series Forecasting Involving Structural Breaks Shijith Kumar PM, PhD Candidate, IE Business School, Calle Maria De Molina, 12, Madrid, 280006, Spain, Matthias Seifert, Yun Shin Lee Forecast combination literature emphasizes the need for detailed studies on aggregation and trimming. We introduce simple trimming rules to aggregate judgmental time series forecasts with structural breaks. While the extant literature explores aggregation of forecasts in stable environments, we focus on environments characterized by fundamental regime shifts. In an empirical study, we find that forecasters are sensitive to structural shifts in time series and are systematically biased depending on the direction of these shifts, making static trimming less applicable. We propose trimming approaches to factor in these biases to aggregate opinion pools under such unstable environments. 3 - Aggregating Information from a Single Set of Predictions Ville Satopaa, INSEAD, 140 Avenue Daumesnil, Paris, 75012, France Even though aggregating multiple predictions typically outperforms the average individual prediction, there is no consensus about the right way to do this. Optimally, an aggregator would use all the information in the predictions. Aggregation is particularly challenging when there is only one prediction per forecaster. In this work, we develop methodology for such a ``one-shot’’ environment. Our aggregator relies on Bayesian statistics and produces a posterior distribution of the consensus aggregate. We illustrate the methodology on real- world and synthetic forecasting data. 4 - Sport Obermeyer Revisited: From Points to Probabilities Asa Palley, Indiana University, 1275 E. 10th St, Bloomington, IN, 47405, United States, Casey Lichtendahl, Yael S. Grushka-Cockayne We develop a procedure that a decision maker can use to estimate a predictive distribution for a variable of interest using only a single point estimate from each of a number of different experts, without having to elicit complete distributions. Given past collections of judgments and realizations, we propose a Bayesian method that can be used to estimate a probability distribution for the variable of interest using the mean and variance of the new collection of judgments. We examine the accuracy of the procedure with sets of real judgments about economic variables in the U.S. and Europe, finding that it provides comparable performance to the linear opinion pool of subjective probability distributions. n MB41 North Bldg 226C Behavioral Decision Analysis with Mortality and Health Outcomes Sponsored: Decision Analysis Sponsored Session Chair: Jeffery L. Guyse, California State Polytechnic University- Pomona, Pomona, CA, 91768, United States Co-Chair: L. Robin Keller, University of California, Irvine, University of California-Irvine, Irvine, CA, 92697-3125, United States 1 - Valuing Sequences of Lives Lost or Saved Over Time: Preference for Uniform Sequences Jeffery L. Guyse, Professor, California State Polytechnic University- Pomona, 3801 West Temple Avenue, Pomona, CA, 91768, United States, L. Robin Keller, Candice Huynh We present our within-subject survey using subjective ratings for sequences of lives lost or saved over time, with factors embedded for anomalies. The prediction results for the standard discounting model (SDM) are analyzed. A model by Loewenstein & Prelec (L&P) for valuation of sequences was then fit to the survey data and compared to the best fits of the SDM. In all cases, the L&P model performed better than the SDM at predicting the individual normalized ratings for these sequences. We conclude that preferences for uniform sequences should be considered in policy making, rather than presuming people have a preference for declining sequences of mortality outcomes.

n MB39 North Bldg 226A Multi-armed Bandits and Reinforcement Learning Sponsored: Applied Probability Sponsored Session Chair: Shipra Agrawal, Columbia University, New York, NY, 10027, United States 1 - Deep Exploration in Reinforcement Learning via Randomized Value Functions Daniel Russo, Columbia University We propose the use of randomized value functions to guide systematic exploration in reinforcement learning. This offers an elegant means for synthesizing statistically and computationally efficient exploration with common RL algorithms like temporal-difference learning, which produce parameterized value function estimates. We present several RL algorithms that leverage randomized value functions and demonstrate their efficacy through computational studies. We also prove a regret bound that establishes statistical efficiency with a tabular representation. 2 - Partition Identification using Multi-armed Bandit Methods with Applications to Financial Risk Sandeep Juneja, Tata Institute of Fundamental Research, Homi Bhabha Road, Mumbai, Maharashtra, 400005, India, Subhashini Krishnasamy Given a vector of probability distributions, each of which can be sampled independently, we identify the partition to which this vector belongs from a finitely partitioned universe using Bandit methods. We develop complexity bounds on expected number of samples needed for correct identification with high probability. Distributions are parametrized, and partitions primarily correspond to these parameters lying in half spaces, or more general convex sets or their complements. We propose algorithms that can match the lower bounds asymptotically with decreasing probability of error. Few applications associated with nested simulation and its applications to finance are discussed. 3 - Bandits with Delayed Aggregated Anonymous Feedback Shipra Agrawal, Columbia University, Industrial Engnr and OR, 423 S. W. Mudd Building, New York, NY, 10027, United States, Randy Jia We study a variant of the stochastic K-armed bandit problem, which we call “bandits with delayed, aggregated anonymous feedback”. In this problem, when the player pulls an arm, a reward is generated, however it is not immediately observed. Instead, at the end of each round the player observes only the sum of a number of previously generated rewards which happen to arrive in the given round due to stochastic delays. We provide an algorithm that matches the worst case regret of the non-anonymous problem exactly when the delays are bounded, and up to logarithmic factors or an additive variance term for unbounded delays. 4 - Near Optimal Policies for Dynamic Assortment Planning under MNL Models Xi Chen, New York University, 44 W. 4th St, NYU KMC Room 8-50, New York, NY, 10012, United States, Yining Wang, Yuan Zhou In this talk, we consider the dynamic assortment selection problem under an uncapacitated multinomial-logit (MNL) model. Since all the utility parameters of customers are unknown, the seller needs to simultaneously learn customers’ choice behavior and make dynamic decisions on assortments based on the current knowledge. By carefully analyzing a revenue potential function, we proposed an efficient trisection algorithm that achieves an item-independent regret bound of O(\sqrt{T\log\log T}), which matches information theoretical lower bounds up to iterated logarithmic terms. We also provide experimental results to demonstrate the effectiveness of the proposed method. n MB40 North Bldg 226B Judgements and Forecasts Sponsored: Decision Analysis Sponsored Session Chair: Asa Palley, Indiana University, Bloomington, IN, 47405, United States 1 - Incentive-Compatible Forecasting Competitions Jens Witkowski, Frankfurt School of Finance & Management, Adickesallee 32-34, Frankfurt, 60322, Germany, Rupert Freeman, Jennifer W. Vaughan, David M. Pennock, Andreas Krause We consider the design of forecasting competitions in which multiple forecasters make predictions about one or more independent events and compete for a single prize. Our primary objective is to incentivize forecasters to report truthfully, so that forecasts are informative and forecasters need not spend any cognitive effort

162