Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TB09

3 - Revenue Management versus Machine Learning in Product Assortment: Evidence from Large Field Experiments on Alibaba Dennis Zhang, University City, MO, 63124, United States, Jacob Feldman We developed a personalized multinomial logit model to find the optimal set of products to display for each incoming customer. We deployed our model in Alibaba’s online recommendation system and compared its performance against the state-of-art machine learning approach used in Alibaba by conducting a large- scale field experiments over more than 40 million customers. Our results show that, despite its low performance in prediction, the MNL-based approach generates higher revenue per customer. 4 - Clearing Matching Markets Efficiently: Informative Signals Andmatch Recommendations Itai Ashlagi, Stanford University, Huang Engineering Center, 475 Via Ortega, Stanford, CA, 94305, United States We study how to reduce congestion in two-sided matching markets with private preferences. We showthat when the unobservable component of agent preferences satisfy certain natural assumptions, it is possible to recommend potential matches and encourage informative signals such that 1)the market reaches an equilibrium outcome, 2) the overall communication overhead is small. The main idea is to only recommend partners whom the agent has a non- negligible chance of both liking and being liked by, as estimated by the observable component of preferences and prior expression of interest by agents on the other side based on the unobservable component of their preferences. n TB11 North Bldg 125B iFORM SIG Sponsored: Manufacturing & Service Oper Mgmt/iFORM Sponsored Session Chair: Shawn Mankad, Cornell University, Cornell University 1 - Credit Risk Propagation Along Supply Chains Jing Wu, City University of Hong Kong, Department of Management Sciences, Hong Kong, Senay Agca, Volodymyr O. Babich, John R. Birge We find that credit risk propagates through multiple supply chain tiers for both positive and negative credit shocks. Specifically, we show sizeable rating and industry-adjusted CDS spread changes of 44-71 bps at the first tier, continuing for 2nd and 3rd tier supply chain partners. Credit risk contagion is more pronounced when supply chain partners are covered by the same analysts. The contagion disappears with inactive supply chain links. The contagion is magnified with longer-term supply-chain relations, trade credit, sales contribution, differentiated products, and customer leverage, while it is moderated when a customer is Starting from the 1990s, banks begin to adopt the Risk Adjusted Return On Capital (RAROC) criterion in loan pricing. In this paper, we study how this criterion influence the bank’s decision on inventory financing, as well as the supply chain implication of this pricing mechanism. 3 - Systemic Risk and Bank Holdings Shawn Mankad, Cornell University The recent financial crisis has focused attention on identifying and measuring systemic risk. We propose a novel approach to estimate the portfolio composition of banks as function of daily interbank trades and stock returns. Our approach estimates bank asset holdings at higher frequencies allowing us to derive precise estimates of (i) portfolio concentration within each bankùa measure of diversificationùand (ii) common holdings across banksùa measure of market susceptibility to propagating shocks. We find evidence that systemic risk measures derived from our approach lead, in a forecasting sense, several commonly used systemic risk indicators. 4 - Regularized Estimation of Factor-augmented Autoregressive Models George Michailidis, University of Florida, UF Informatics Institute, 432 Newell Drive, Gainesville, FL, 32611-8545, United States We consider the factor-augmented vector autoregressive (FAVAR) model in the high dimensions, where the large observed block and the latent factor block jointly follow a vector autoregressive (VAR) model, with an additional information series linked to both of them contemporaneously through a linear model. We address model identifiability and estimation issues and examine the performance of the model through simulations. Further, we apply the model to an economic dataset involving commodity prices and macroeconomic sequences to investigate interlinkages amongst the former. investment grade or has more inventory. 2 - Inventory Financing Under Raroc Yuxuan Zhang, London Business School, London, United Kingdom, Simin Huang, S. Alex Yang

n TB09 North Bldg 124B Nonconvex Optimization in Machine Learning Sponsored: Optimization/Nonlinear Programming Sponsored Session Chair: Ruoyu Sun, UIUC, Minneapolis, MN, 55414, United States 1 - Gradient Descent with Random Initialization: Fast Global Convergence for Solving Quadratic Systems of Equations Yuxin Chen, Princeton, Yuejie Chi, Jianqing Fan, Yuejie Chi This talk considers the problem of solving systems of quadratic equations. We investigate the efficiency of gradient descent designed for the nonconvex least squares problem. We prove that under Gaussian designs, gradient descent ù when randomly initialized ù yields an -accurate solution in a logarithmic number of iterations given nearly minimal samples, thus achieving near-optimal computational and sample complexities at once. This provides the first global convergence guarantee concerning vanilla gradient descent for phase retrieval, without the need of (i) carefully-designed initialization, (ii) sample splitting, or Agents trained with reinforcement learning (RL) are usually reactive: they take actions from the observations, perceive new observations and repeat. However, human doesn’t behave this way. They plan by predicting what would be possible futures if some actions were taken, and pick the actions that lead to the best consequence. This involves a strong modeling of the world and its dynamics, which is non-trivial to obtain even for simulated situations such as games. In this talk, we introduce our recent works about how planning is done in RL, and how RL can help perform better planning 3 - Understanding Landscape of Neural Networks for Binary Classification Ruoyu Sun, UIUC, IL, United States We study the landscape of neural networks for binary classification and provide conditions under which the training error is zero at all local minima. We provide a rather comprehensive study on necessary conditions as well as sufficient conditions. For example, we show that while ReLU and sigmoid activations lead to bad local minima, strictly increasing and convex activation functions can eliminate bad local minima. Other conditions include: the neural network should either be single-layered or is multi-layered with a shortcut-like connection, and the surrogate loss function should be a smooth version of hinge loss. n TB10 North Bldg 125A Market Design Sponsored: Manufacturing & Service Oper Mgmt Sponsored Session Chair: Gabriel Weintraub, Stanford Graduate School of Business, Stanford, CA, 94304, United States 1 - How (Not) to Allocate Affordable Housing Peng Shi, University of Southern California, USC Marshall School of Business, BRI 303D, 3670 Trousdale Pkwy, Los Angeles, CA, 90089, United States, Nicholas A. Arnosti Motivated by affordable housing allocation in NYC, we study a stylized model in which items are dynamically assigned to waiting agents. The planner has two goals—to match each agent with the most fitting item, and to target the agents with the highest need. We find that the current practice of independent lotteries encourages agents to enter many lotteries, resulting in inefficient matching. The matching can be improved by limiting the number of lottery entries, but this can result in worse targeting. We prove that this tradeoff between matching and targeting cannot be avoided as long as the levels of need are private knowledge, and give guidance on which to prioritize to maximize utilitarian welfare. 2 - Designing Framework Agreements: A Field Study in Government Procurement Marcelo Olivares, Associate Professor, Universidad de Chile, Santiago, Chile, Gabriel Weintraub, Daniela Saban, Eduardo Lara Framework Agreements (FA) is a commonly used procurement mechanism by governments and large organizations. It is based on an auction-type design to select an assortment of products from multiple suppliers with posted prices, allowing some flexibility and variety to purchasing units. The design of FA requires balancing competition to enter the market with the variety offered inside the market. This paper conducts an empirical study of the FAs used by the Chilean government, identifying inefficiencies in the procurement market, providing improvements to design and conducting a field study to measure the actual effectiveness of the new implemented design. (iii) sophisticated saddle-point escaping schemes. 2 - Planning for RL and RL for Planning Yuandong Tian, Facebook, Menlo Park, CA, USA.

278