Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SD25

4 - An Approximate Dynamic Programming Approach to Dynamic Pricing for Network Revenue Management Jiannan Ke, Shanghai Jiao Tong University, 1954 Huashan Road, Zhongyuan 111, Shanghai, 200030, China, Dan Zhang, Huan Zheng We propose an approximate dynamic programming approach to the dynamic pricing problem for network revenue management. The approximate linear program (ALP) are semi-infinite linear programs and can be solved to any desired accuracy with a column generation algorithm. For the affine approximation under a linear independent demand model, we show that the ALP can be reformulated as a compact second order cone program (SOCP). The size of the SOCP formulation is linear in the number of resources, products, and periods. Numerical experiments show that solving the SOCP formulation is orders of magnitude faster and the ensuing pricing policies perform well.

n SD24 North Bldg 131B eBusiness on Online Two-sided Market Sponsored: EBusiness Sponsored Session Chair: Jinyang Zheng, Purdue University, West lafayette, IN, 47906, United States 1 - What Goes Around Comes Around: A Structural Matching Model of Peer-to-Peer Lending Jinyang Zheng, Purdue University, West lafayette, IN, 47906, United States, Yang Jiang, Xiangbin Yan, Yong Tan Our work is the first empirical study that investigates the matching mechanism of online P2P lending. In this paper, we use data on transaction records from a large lending site, and apply a structural matching model to examine the determinants of a lender-borrower match formation. Our results show that the matching criteria differ across heterogeneous lenders and borrowers, depending on the interaction characteristics between the two sides. We provide evidence that a larger match value is associated with a better loan performance, which validates the effectiveness of the estimated matching pattern. Our findings shed lights on how to improve the matching efficiency of online two-sided markets. 2 - Does Size Matter? The Effect of Sampling Size in Online Physical Good Sampling Zibo Liu, University of Washington, Seattle, WA, 98195-5832, United States, Zhijie Lin, Ying Zhang, Yong Tan Despite the popular use of product sampling as a promotional strategy by retailers, existing research has only studied offline sampling of physical goods and online sampling of information goods, but overlooked online sampling of physical goods. We argue that, in the context of online sampling of physical goods, sampling size serves as a signal of product quality to positively influence product sales, and this effect would vary across product type. We collected a panel-level data set from Taobao.com for empirical analysis. We find evidence for the existence of the effect of sampling size in online sampling of physical goods, and find that the effect indeed vary across products. 3 - Consumer Decisions in Payment-based Knowledge Sharing Communities: The Analysis of Zhihu Live Zhen Fang, University of Washington, Seattle, WA, 98195-5832, United States, Guannan Liu, Junjie Wu, Yong Tan Knowledge monetization is becoming a new trend in knowledge sharing communities. Experts on platforms as Zhihu Live give payment-based lectures to audience in the community. This research empirically investigates the impact of different factors on the consumer decisions, including the price, communications, online reviews and the reputation of speakers in Q&A community, Zhihu. We conclude that lower price, more communication and higher reputation of the speaker improve the popularity of Lives. n SD25 North Bldg 131C Service Science Topics in the IT Industry Sponsored: Service Science Sponsored Session Chair: Aly Megahed, IBM Research - Almaden, San Jose, CA, 95123, United States 1 - Will they Sign? Predicting the Contract Pipeline with Structured and Unstructured Data Aly Megahed, IBM Research - Almaden, San Jose, CA, 95123, United States, Hamid Reza Motahari Nezhad, Paul R. Messinger Suppliers of products and services often compete for highly-valued contracts in lengthy tender processes. Prediction of success for the various prospects in the pipeline is vital to managing a company’s work flow. Toward this end, we develop a contract prediction model based on two kinds of data: structured deal attributes and unstructured text of seller comments. We show real-world numerical results that illustrate the effectiveness of our models. We also present managerial implications and insights of implementing this work for a real IT firm. 2 - Analytics Driven Business Travel Management Solution Pawan Chowdhary, IBM Research, 650 Harry Road, E3-238, San Jose, CA, 95120, United States, Raphael Arar, Guangjie Ren, Sunhwan Lee Business travel is essential to meet customers and for the company to grow. Mid to large enterprise has business travel as one of the largest spend items but still not managed efficiently. IBM Travel Manager is a solution that is catered for the Travel Program Manager to effectively manage travel spend and leverage analytics to get spend insights and savings opportunities. In the presentation we will go few such analytics and touch over cognitive aspect that will drive the travel manager to act upon opportunities pro-actively.

n SD23 North Bldg 131A Statistical Methods in Finance

Sponsored: Finance Sponsored Session Chair: Markus Pelger, Stanford University, Stanford, CA, 94305, United States 1 - Efficient Computational Methods for Distributionally Robust Optimization with Martingale Constraints Jose Blanchet, Columbia University, 500 West 120th Street, 340 Mudd Building, New York, NY, 10027, United States We study efficient distributionally robust optimization methods under martingale constrains. These problems arise in several OR applications, including. 2 - Deep Learning Models of High Frequency Financial Data Justin Sirignano, University of Illinois at Urbana-Champaign, 3 Gibbs Court, Irvine, CA, 92617, United States Using a Deep Learning approach applied to a large dataset of high frequency financial data, we find evidence for a universal and stationary price formation mechanism relating the supply and demand for a stock, as revealed through the order book, to price dynamics. We build a ‘universal price formation model’ which demonstrates stable accuracy across a wide range of stocks from different sectors and for long time periods. The universal model, trained on data from all stocks, outperforms asset-specific linear and nonlinear models trained on time series of any given stock. This shows that the universal nature of price formation weighs in favor of pooling together financial data from various stocks, rather than designing asset- or sector-specific models as commonly done. We also find that price formation has path-dependence over long periods of time (‘long memory’). Joint work with Rama Cont. 3 - Statistical Inference using Neural-Networks Enguerrand Horel, PhD Student, Stanford University, Stanford, CA, 94305, United States, Kay Giesecke Although neural networks can provide highly accurate predictions, they are often considered as opaque “black boxes”. The difficulty of interpreting the predictions of a neural network often prevents their use in financial practice, where regulators and auditors often insist on model explainability. In this project, we formulate the objectives of interpretability and define interpretability for neural networks. Then, by considering neural networks as nonparametric models, we show how to use nonparametric variable significance tests to assess the importance of covariates. Tests using mortgage data illustrate our results. 4 - Interpretable Proximate Factors for Large Dimensions Ruoxuan Xiong, Stanford University, 312 Huang Engineering Center, Stanford, CA, 94305, United States, Markus Pelger This paper approximates latent statistical factors with sparse and easy-to-interpret proximate factors. Latent factors in a large-dimensional factor model can be estimated by principal component analysis, but are usually hard to interpret. By shrinking factor weights, we obtain proximate factors that are easier to interpret. We show that proximate factors consisting of 5-10% of the observations with the largest absolute loadings are usually sufficient to almost perfectly replicate the population factors without assuming a sparse structure. We derive lower bounds for the generalized correlation between proximate and population factors based on extreme value theory.

99