2016 INFORMS Annual Meeting Program

SA38

INFORMS Nashville – 2016

3 - Inventory Management In An Omnichannel Environment Yong-Pin Zhou, Professor, Foster School of Business, University of Washington, Seattle, Seattle, WA, 98195-3226, United States, yongpin@uw.edu, Elnaz Jalilipour Alishah We consider a single newsvendor-type product that is sold both online and offline. We present two models where each channel is used as a backup for the other channel, and derive structural and qualitative results on effective inventory management policies. Specifically, we consider inventory positioning, inventory level, and real-time inventory rationing decisions. When it is possible to shift some customer demand using discounts, we also investigate the level of discount and customer reaction. 4 - The Omni-channel Fulfillment Dilemma

4 - Institutional Design And The Creative Process: An Experimental Study Lakshminarayana Nittala, University of California San Diego, La Jolla, CA, 92037, United States, lnittala@ucsd.edu Sanjiv Erat, Viswanathan Krishnan The process of Innovation often takes the form of problem solving and requires creative insights for achieving success. In an experimental setting we use tasks that are representative of such problems and study the role of institutional design on the creative output and the underlying search process. SA39 207A-MCC A/B Testing, Experiments, and Learning Sponsored: Applied Probability Sponsored Session Chair: Ramesh Johari, Stanford University, Stanford, CA, United States, ramesh.johari@stanford.edu Co-Chair: David Walsh, Stanford University, Department of Statistics, Stanford, CA, 94305, United States, dwalsh@stanford.edu 1 - Using Simulation To Improve Statistical Power In Switchback Experiments At Uber Peter Frazier, Cornell University, Ithaca, NY, 14850, United States, pf98@cornell.edu We consider A/B testing of systemic changes with time-varying effects, such as changes to the algorithm used to dispatch cars at Uber. Testing such changes is made difficult by correlations in outcomes across dispatches, and by seasonal and autocorrelated random variation in riders’ demand for trips. One standard A/B testing method is a switchback experiment, which applies the treatment and control on alternating days over two weeks. We show how to combine simulation-based predictions of a change’s effects with data from a switchback experiment to improve statistical power, and make analysis robust to missing data. 2 - A/B Testing In A Changing World David Walsh, Stanford University, dwalsh@stanford.edu The purpose of A/B testing is to let any technology company iterate on its products quickly and stay ahead of a rapidly changing market. Given this dynamic context, it is odd that the existing statistical approaches to A/B testing view the environment as static. The outcome: inferences that do not generalize beyond the life of the experiment, which then lead to actions that perform substantially worse than expected. I present new methodology that anticipates temporal variation, generating the right inferences to support dynamic product optimization. 3 - Simple Bayesian Algorithms For Identifying The Best Arm In A Multi-armed Bandit Daniel Russo, Northwestern University, Evanston, IL, United States, Dan.Joseph.Russo@gmail.com This talk considers the optimal adaptive allocation of measurement effort for identifying the best among a finite set of options or designs. An experimenter sequentially chooses designs to measure and observes noisy signals of their quality with the goal of confidently identifying the best design after a small number of measurements. I propose three simple Bayesian algorithms for adaptively allocating measurement effort. Each is shown to have strong performance in numerical experiments, and a unified analysis establishes each Vivek Farias, Massachusetts Institute of Technology, Cambridge, MA, United States, vivekf@mit.edu, Ciamac Cyrus Moallemi We consider the problem of sequential A-B testing when the impact of a treatment is marred by a large number of covariates. Our main contribution is a tractable algorithm for the online allocation of test subjects to either treatment with the goal of maximizing the efficiency of the estimates treatment effect under a linear model, which due to a surprising state space collapse, reduces to solving a low dimensional dynamic program. Our approach is robust and covers many variations of the problem, including cases where there are budget constraints on individual treatments, where the number of trials is to be endogenously decided, and where the objective is to balance a tradeoff between efficiency and bias. satisfies a strong asymptotic optimality property. 4 - Sequential A-B Testing With Constraints

Santiago Gallino, Dartmouth College, Dartmouth, NH, United States, santiago.gallino@tuck.dartmouth.edu, Antonio Moreno-Garcia, Robert P. Rooderkerk

Using transaction level data from a multi-channel retailer we explore how customer interact with the retailer over time. We study the underlying patters of customer’s interactions and discuss the implications for retailers that have both an online and a brick and mortar presence.

SA38 206A-MCC Product Development and Competition Invited: New Product Development Invited Session Chair: Morvarid Rahmani, Georgia Institute of Technology, Atlanta, Atlanta, GA, 30308, United States, morvarid.rahmani@scheller.gatech.edu Co-Chair: Karthik Ramachandran, Georgia Institute of Technology, Atlanta, GA, 30308, United States, karthik.ramachandran@scheller.gatech.edu 1 - Knowledge Search In Mobile App Development Nilam Kaushik, PhD Candidate, UCL School of Management, London, United Kingdom, uceikau@ucl.ac.uk, Bilal Gokpinar The process of search, identification, and acquisition of new knowledge is essential for the success of new products. We explore how firms search for ideas in sequential product development through the highly competitive and dynamic setting of mobile application development. Using novel text-mining techniques, we derive measures of similarity between a focal app’s update and updates made by competitor apps and study performance implications thereof. We also explore the performance implications of the distance of an app’s update with respect to its past updates. 2 - Decision Options At Project Gate Reviews: Beyond The Go/ Kill Model Alison Olechowski, Massachusetts Institute of Technology, Cambridge, MA, United States, alisono@mit.edu Steven D Eppinger, Nitin Joglekar Most current academic models of project gate reviews represent the gate decision as a simple choice between go and kill. In reality, product developers often reach the gate with incomplete or unacceptable deliverables, and managers consider more than just the kill option if a go is not appropriate. We have created a simple model which adds options of: waiver, waiver with re-review, delay and switch to back-up plan. This decision tree model compares the value of this more realistic set of options based on costs, payoffs and confidences. We demonstrate the application of this model to complex product development gate decisions via multiple case examples from industry. 3 - Capacity Investment For Product Upgrades Under Competition Ram Bala, Santa Clara University, ram.bala@gmail.com, Milind Sohoni, Sumit Kunnumkal Firms often introduce a vertical line extension of an existing product to consolidate their market position after loss of monopoly status. However, introducing a line extension is fraught with uncertainty as it may fail to be technically feasible as originally intended. We analyze a two stage competitive game between an incumbent and an entrant where the firms make capacity investment decisions before uncertainty resolution and then set production quantities. We uncover conditions on innovation level which determine whether the incumbent will continue to offer the existing product once the new product succeeds. We also determine the innovation level beyond which competitive entry is deterred.

30