Modeling User Click Behavior in Sponsored Search
|
|
- Ralf Wells
- 6 years ago
- Views:
Transcription
1 Modeling User Click Behavior in Sponsored Search Vibhanshu Abhishek, Peter S. Fader, Kartik Hosanagar The Wharton School, University of Pennsylvania, Philadelpha, PA 1914, USA {vabhi, faderp, Abstract There has been significant recent interest in studying consumer behavior in sponsored search environments. Researchers have typically used aggregate data provided to advertisers by search engines to model consumer behavior. The data typically catalogue advertiser s bid, average position and total impressions, clicks and cost on a daily basis for keywords in the advertisers sponsored search campaign. Random utility models have been proposed and their parameters estimated using logistic regression. We show that estimating these models on aggregated data can lead to biased estimates, specifically the impact of ad position on consumer click propensity is often underestimated. L étude du comportement des consommateurs dans des environnements de recherche sponsorisée a été objet d un grand intérêt récent. Les investigateurs ont généralement utilisé des données agrégées fournies aux annonceurs par les moteurs de recherche dans la construction des modèles de comportement des consommateurs. Les modèles d utilité aléatoire ont été proposées et leurs paramètres estimés à l aide de la régression logistique. Nous montrons que l estimation de ces modèles sur des données agrégées peut conduire à des estimations biaisées, et plus précisément l impact de la position des annonces sur la propension des consommateurs à cliquer l annonce est souvent sous-estimé. 1 Introduction Sponsored search is the fastest growing form of advertising on the Internet. A number of factors have contributed to this growth. The ads tend to be highly targeted and offer a higher return on investment for advertisers compared to other marketing methods. In addition the large audience it offers has led to a wide-spread adoption of search engine advertising. When a user issues a query on the search engine, sponsored results are displayed alongside organic search results. The organic search results are links relevant to the query and are ranked in order of their relevance. The sponsored results are ads submitted by advertisers. The advertisers submit bids for keywords that are relevant to them, along with these ads. When a user enters a query, the search engine identifies the advertisers bidding on keywords closely related to the query and uses data on bids and ad quality/performance to rank order the ads that appear in the list of sponsored results. The most widely used pricing model is the pay per click model, in which the advertiser pays only when a user clicks on his ad. The advertiser s cost per click or cpc is determined using a generalized second price auction [7], i.e. whenever a user clicks on an ad in position, the advertiser pays an amount equal to the minimum bid needed to secure that position. One of the attractive features of sponsored search is that it is a highly measurable form of advertising. Data on consumer click and purchase patterns have been used to study consumer behavior and advertiser strategies. Several researchers have built random utility models to study the effect of ad position, keyword length, presence or absence of brand name, etc. on the clickthrough rate (ctr) of the ad. The parameters of these models have been estimated from aggregated data that catalogue advertiser s bid, average position and 1
2 total impressions, clicks and cost on a daily basis for keywords in the advertisers sponsored search campaign. In this paper we show that applying standard models (e.g., logistic regression) on such aggregated datasets can lead to biased estimation of the parameters of a random utility model. Using a simple model we show that the effect of position on ctr is underestimated when the coefficients are estimated using the summarized data. 2 Related work There has been a lot of recent interest in trying to understand the factors driving keyword performance in sponsored search [3, 8, 1]. These studies typically propose individual keyword-level models. However, clicks and conversions are often very sparse in most sponsored search data, making it hard to estimate individual-level models. Rutz and Bucklin [8] propose a hierarchical bayesian model to study the conversion performance of individual keywords. They show that the model adequately addresses the sparse data problem while accounting for keyword heterogeneity. In another related paper Rutz and Bucklin [9] study the performance of branded versus generic keywords and identify spill over effects of generic keywords using the Nerlove-Arrow goodwill model. Ghose and Yang [4] use hierarchical bayesian models to quantify the impact of ad position, keyword length and type and the quality of the landing page on consumer conversion behavior and the advertiser s cpc. Ghose and Yang also show that the advertisers are not bidding optimally to maximize their profits. Recent work by Agarwal et. al [1] shows that although the ctr decreases with position, the conversion rate often increases and then decreases. They show that the topmost position is not necessarily the revenue maximizing position. The rank of an ad varies across impressions during a day, but since these use aggregate data, their models assume the mean rank during the day is the actual rank. This aggregation of data can lead to potential biases in the estimation of parameters and ultimately affect the conclusions from these studies. The problem of aggregation bias is not new and has been addressed earlier in some detail. Kelejian [6] discusses why, under certain conditions, aggregation bias might occur and proposed a test for the existence of this bias. Logistic regression have been extensively used in Marketing where hedonistic utility models have been estimated using aggregate scanner data. Allenby and Rossi [2] present an analytical proof for the non existence of aggregation bias in nested logit models when the micro level consumer behavior is assumed to be linear in nature and the products are close substitutes of each other. This paper presents a conclusive proof for the existence of the bias in the setting of sponsored search. In addition it clearly outlines the direction of the aggregation bias, which we believe has been shown for the first time. 3 Bias due to Aggregate Data In this section we explore the estimation bias introduced due to the aggregation of data. First, we analytically show that the logistic regression gives biased estimates on aggregated data. Then we try to derive insights about the direction of the bias. Finally, we present a simulation study to illustrate the results and identify the conditions under which the bias is negligible versus significant. 3.1 Analytical Proof of the Bias Logit utility model The logit utility model has been extensively used in marketing and economics to model consumer choice behavior. Suppose the consumer utilility is modeled as follows: u(pos) = β β 1 pos ε. (1) If ε has a logistic distribution the probability of clicking on an ad at position pos is given by p(y = 1 pos; β, β 1 ) = Λ(β β 1 pos), (2) where Λ(β β 1 pos) = e β β 1 pos /[1 e β β 1 pos ]. In this simple model we ignore other ad characteristics
3 to focus our attention on the impact of ad position on click through rate. The simple model allows us to clearly identify the existence and direction of the bias. The log likelihood of observing a data-set with a total of I ad impressions is given by I LL(β data) = y i log Λ(x iβ) (1 y i ) log(1 Λ(x iβ)), (3) i=1 where x i = (1 pos i ), pos i is the position of the ad on the i th impression and y i is equal to 1 if the i th impression resulted in a click and zero otherwise. The maximum likelihood estimate ˆβ complete for the complete data can be computed by ˆβ complete = arg max y i log Λ(x iβ) (1 y i ) log(1 Λ(x iβ)). (4) β We can show that ˆβ complete satisfies obsctr = C I = 1 I t I Λ(x ˆβ i complete ), where C = i=1 where x i is the transpose of the column vector x i Proposition 3.1 ˆβ complete is a consistent estimator of β. I y i, (5) Proof Since {y i, x i } constitute an ergodic stationary sequence and conditions of identification and dominance are satisfied, the logit model gives consistent estimates on the complete data (Hayashi [5] Proposition 7.6) In case of summary data over D days the LL of observing a given aggregated dataset is LL(β data) = ( ) nd log c c d log Λ(x d β) (n d c d ) log(1 Λ(x dβ)), (6) d d where x d = (1 pos d ) T, pos d is the average position on the d th day, n d is the total number of impressions and c d is the total number of clicks on day d.the estimate for summary data ˆβ summary is computed by D ( ) nd ˆβ summary = arg max log c d log Λ(x d β) (n d c d ) log(1 Λ(x dβ)), (7) β which satisfies c d obsctr = C I = D i=1 n d I Λ(x d ˆβ summary ). (8) For the sake of discussion we introduce the following notation. Let V i 1 be a random variable denoting the position in the i th impression. We assume that V i is independently and identically distributed and has a distribution given by F V () which is assumed to be constant during the period of observation. The average position for the day given that there were n queries in the day is U n = V 1 V 2... V n (9) n The distribution of U n is F Un () which is given by n and the distribution of V. An important relationship between U n and V is as follows. Lemma 3.2 U n is less than V in convex order. U n cx V (1) 1 pos i is the realization of V i.
4 X is less than Y in convex order if E[f(X)] E[f(Y )] for all real convex functions f such that the expectation exists. The proof of this lemma has been omitted due to space constraints. Now we present an important result of this paper. Proposition 3.3 ˆβ summary is a biased estimator of β. Proof We first prove this general result using a special case when n d = n for all d = 1,..., D. Substituting U n in place of pos d, Equation (8) can be simplified as D n obsctr = I Λ(x ˆβ d summary ) = 1 D e ˆβ,summary ˆβ 1,summary U n D 1 e ˆβ,summary ˆβ. 1,summary U n Assuming D is large we can apply Chebyshev s law of large numbers to get e ˆβ,summary ˆβ 1,summary u obsctr = 1 e ˆβ,summary ˆβ df U 1,summary u n (u). (11) Similarly V is substituted in Equation (5) and assuming that I is large we can rewrite the summation as the following integral obsctr = e ˆβ,complete ˆβ 1,complete v 1 e ˆβ,complete ˆβ 1,complete v df V (x). (12) Since the observed ctr is same in both the cases e ˆβ,summary ˆβ 1,summary u 1 e ˆβ,summary ˆβ df e ˆβ,complete ˆβ 1,complete v U 1,summary u n (u) = 1 e ˆβ,complete ˆβ df V (x). (13) 1,complete v Since Equation (13) always holds we will now show that ˆβ complete ˆβ summary and hence ˆβ summary is biased. Since the convex ordering in lemma 3.2 holds and logit is a convex function for the set of parameters considered 2 it follows from the definition of convex ordering that e ˆβ,summary ˆβ 1,summary u 1 e ˆβ,summary ˆβ df U 1,summary u n (u) 1 e ˆβ,complete ˆβ df V (x), (14) 1,complete v if ˆβ complete = ˆβ summary. Since Equation (13) and Equation (14) cannot hold simultaneously and Equation (13) always holds we prove by contradiction that ˆβ complete ˆβ summary e ˆβ,complete ˆβ 1,complete v Now we relax the assumption that n d = n for all d = 1,..., D. Let U = U N represent the average position when N, the number of queries in a day, is random. It can be shown that U cx V which gives us e ˆβ,summary ˆβ 1,summary u 1 e ˆβ,summary ˆβ 1,summary u df U(u) e ˆβ,complete ˆβ 1,complete v 1 e ˆβ,complete ˆβ 1,complete v df V (x). (15) if ˆβ complete = ˆβ summary. As shown earlier this cannot happen hence ˆβ complete ˆβ summary Proposition 3.4 The direction of aggregation bias is such that Λ( ˆβ summary, x) is greater than Λ( ˆβ complete, x) x. Proof Since we know that and by definition of convex order Λ( ˆβ summary, x)df U (x) = Λ( ˆβ complete, x)df V (x) Λ( ˆβ complete, x)df V (x), Λ( ˆβ complete, x)df U (x) 2 We can show that the logit is convex whenever β which is a reasonable assumption given that the probability of the first ad in the list getting clicked is always seen to be significantly less than.5. The result holds even when this condition is relaxed as is shown in the simulation.
5 Λ( ˆβ summary, x) Λ( ˆβ complete, x) 3.2 Illustrative Examples We now turn to simulations to validate the analytical results and identify the cases in which the magnitude of the bias is significant. The inputs to the simulation are the coefficients of the logit model (β, β 1 ) and the mean arrival rate λ for a Poisson process that determines the number of number of queries in a day. The consumer s decision to click on the ad conditional on its position is simulated for each query using the random utility model specified earlier 3. The complete data record the position and the binary click decision for all queries. In addition, the daily total number of impressions, clicks and the mean position for our ad are computed and recorded in the aggregate data. 1 different runs are generated for every set of β, β 1 and λ. The graph in Figure 1 shows the ctr v/s position for summary and complete data for one particular run ctr..2.4 o summary ctr complete ctr pos Figure 1: ctr v/s position with β =, β 1 =.75, λ = 1. Crosses are the ctr observed at each position in the complete data set and the dotted line is the logit estimated on the complete data. The actual underlying logit function is close to the dotted line and has been omitted for clarity. The dots represent the observed ctr for various daily mean positions during the observation period and the dashed line is the logit function estimated on the summary data. It intuitively follows from Jensen s inequality that averaged ctr is greater than the actual ctr. Since the regression estimates the parameters on the averaged observations the estimated ctr is greater than the actual logit model. β,actual β 1,actual β,complete β 1,complete β,summary β 1,summary Table 1: Estimated parameters of logit model on summary data. The reported values are the mean of the estimated coefficients for every set of parameters. implies that the mean lies within the 95% confidence interval about the actual parameter on a one sample t-test We report the results for several βs in Table 1 to illustrate the nature of the bias for different values of β 4. We observe that as the effect of position on ctr increases the aggregation bias increases. This is due to the increasing convexity of the ctr-position curve. Further, the effect of position is underestimated in all cases. When the decay in ctr with position is not very steep, the aggregation bias is small because ad position 3 The decision is a binary variable indication whether the user clicked our ad or not 4 The results are robust to variation in λ.assuming that λ is sufficiently high, the law of large numbers comes into play.
6 has little impact on the number of clicks. However as position becomes more important and ctr decreases significantly the bias increases. In additional simulations with other covariates in a richer logit model, we find that this leads to an overestimation of coefficients of regressors that contribute positively to the utility function of the consumer. This implies that some other ad characteristics might have received more credit in prior work than is justified. Simulations show that the results are robust to variations in F V (). 4 Discussion and Conclusion The search engine advertising environment offers rich data on consumer click and conversion behavior. Several recent studies have started analyzing such data to study consumer behavior and inform advertiser strategies. A key contribution of this paper is that it demonstrates that the techniques used in several studies in the literature might not be appropriate given the aggregated nature of the data. We analytically prove the existence of aggregation bias and identify the direction of the bias. We further show that the bias causes the effect of position on ctr to be underestimated. This leads not only to biased estimates for the coefficient of position but also results in flawed estimation of the contribution of various factors on the clickthrough rate. As an extension of this work we are developing a probabilistic model for consumer behavior that accounts for variation in position. Preliminary results show that the parameters of the model can be estimated even from aggregate data. References [1] A. Agarwal, K. Hosanagar, and M. D. Smith. Location, location, location: An analysis of profitability and position in online advertising markets. SSRN elibrary, 28. [2] G. M. Allenby and P. E. Rossi. There is no aggregation bias: Why macro logit models work. Journal of Business and Economic Statistics, pages 1 14, [3] N. Craswell, O. Zoeter, M. Taylor, and B. Ramsey. An experimental comparison of click position-bias models. In WSDMŠ8, 28. [4] A. Ghose and S. Yang. An empirical analysis of search engine advertising: Sponsored search and cross-selling in electronic markets. SSRN elibrary, 27. [5] F. Hayashi. Econometrics. Princeton University Press, 2. [6] H. H. Kelejian. Aggregated heterogeneous depentdent data and the logit model: A suggested approach. Economic Letters, pages , [7] S. Lahaie and D. M. Pennock. Revenue analysis of a family of ranking rules for keyword. In EC7. [8] O. J. Rutz and R. E. Bucklin. A model of individual keyword performance in paid search advertising. SSRN elibrary, 27. [9] O. J. Rutz and R. E. Bucklin. From generic to branded: A model of spillover dynamics in paid search advertising. SSRN elibrary, 28.
Essays in Online Advertising
University of Pennsylvania ScholarlyCommons Publicly Accessible Penn Dissertations 1-1-2012 Essays in Online Advertising Vibhanshu Abhishek University of Pennsylvania, vibs@cmu.edu Follow this and additional
More informationOur MCMC algorithm is based on approach adopted by Rutz and Trusov (2011) and Rutz et al. (2012).
1 ONLINE APPENDIX A MCMC Algorithm Our MCMC algorithm is based on approach adopted by Rutz and Trusov (2011) and Rutz et al. (2012). The model can be written in the hierarchical form: β X,ω,Δ,V,ε ε β,x,ω
More informationDo Organic Results Help or Hurt Sponsored Search Performance?
Do Organic Results Help or Hurt Sponsored Search Performance? Ashish Agarwal Assistant Professor McCombs School of Business University of Texas, Austin Austin, TX 15213 Phone: (512) 471 5814 Email: ashish.agarwal@mccombs.utexas.edu
More informationExtended Abstract Prepared for Submission to WISE 2012 Social Advertising: Does Social Influence Work?
Extended Abstract Prepared for Submission to WISE 2012 Social Advertising: Does Social Influence Work? Ashish Agarwal McCombs School of Business U T Austin ashish.agarwal@mccombs.utexas.edu Kartik Hosanagar
More informationLocation, Location, Location: An Analysis of Profitability of Position in Online Advertising Markets
Carnegie Mellon University Research Showcase @ CMU Heinz College Research Heinz College 5-1-2008 Location, Location, Location: An Analysis of Profitability of Position in Online Advertising Markets Ashish
More informationDo Organic Results Help or Hurt Sponsored Search Performance
University of Pennsylvania ScholarlyCommons Finance Papers Wharton Faculty Research 12-2015 Do Organic Results Help or Hurt Sponsored Search Performance Alekh Agarwal kartik Hosanagar University of Pennsylvania
More informationSponsored Search: Do Organic Results help or hurt the Performance and under what conditions?
Sponsored Search: Do Organic Results help or hurt the Performance and under what conditions? Ashish Agarwal Assistant Professor McCombs School of Business University of Texas, Austin Austin, TX 15213 ashish.agarwal@mccombs.utexas.edu
More informationLocation, Location, Location: An Analysis of Profitability of Position in Online Advertising Markets
ASHISH AGARWAL, KARTIK HOSANAGAR, and MICHAEL D. SMITH The authors evaluate the impact of ad placement on revenues and profits generated from sponsored search. Their approach uses data generated through
More informationWRITTEN PRELIMINARY Ph.D. EXAMINATION. Department of Applied Economics. University of Minnesota. June 16, 2014 MANAGERIAL, FINANCIAL, MARKETING
WRITTEN PRELIMINARY Ph.D. EXAMINATION Department of Applied Economics University of Minnesota June 16, 2014 MANAGERIAL, FINANCIAL, MARKETING AND PRODUCTION ECONOMICS FIELD Instructions: Write your code
More informationOnline Ad Auctions. By Hal R. Varian. Draft: December 25, I describe how search engines sell ad space using an auction.
Online Ad Auctions By Hal R. Varian Draft: December 25, 2008 I describe how search engines sell ad space using an auction. I analyze advertiser behavior in this context using elementary price theory and
More informationField Exam January Labor Economics PLEASE WRITE YOUR ANSWERS FOR EACH PART IN A SEPARATE BOOK.
University of California, Berkeley Department of Economics Field Exam January 2017 Labor Economics There are three parts of the exam. Please answer all three parts. You should plan to spend about one hour
More informationEstimating Discrete Choice Models of Demand. Data
Estimating Discrete Choice Models of Demand Aggregate (market) aggregate (market-level) quantity prices + characteristics (+advertising) distribution of demographics (optional) sample from distribution
More informationUnderstanding UPP. Alternative to Market Definition, B.E. Journal of Theoretical Economics, forthcoming.
Understanding UPP Roy J. Epstein and Daniel L. Rubinfeld Published Version, B.E. Journal of Theoretical Economics: Policies and Perspectives, Volume 10, Issue 1, 2010 Introduction The standard economic
More informationWRITTEN PRELIMINARY Ph.D. EXAMINATION. Department of Applied Economics. January 27, Consumer Behavior and Household Economics.
WRITTEN PRELIMINARY Ph.D. EXAMINATION Department of Applied Economics January 27, 2017 Consumer Behavior and Household Economics Instructions Identify yourself by your code letter, not your name, on each
More informationCapacity Dilemma: Economic Scale Size versus. Demand Fulfillment
Capacity Dilemma: Economic Scale Size versus Demand Fulfillment Chia-Yen Lee (cylee@mail.ncku.edu.tw) Institute of Manufacturing Information and Systems, National Cheng Kung University Abstract A firm
More informationModeling of competition in revenue management Petr Fiala 1
Modeling of competition in revenue management Petr Fiala 1 Abstract. Revenue management (RM) is the art and science of predicting consumer behavior and optimizing price and product availability to maximize
More informationTesting the Signaling Theory of Advertising: Evidence from Search Advertisements
Testing the Signaling Theory of Advertising: Evidence from Search Advertisements Taylor Bentley, Tat Y. Chan and Young-Hoon Park January 2015 Taylor Bentley is a PhD Candidate in Marketing at the Olin
More informationProgress Report: Predicting Which Recommended Content Users Click Stanley Jacob, Lingjie Kong
Progress Report: Predicting Which Recommended Content Users Click Stanley Jacob, Lingjie Kong Machine learning models can be used to predict which recommended content users will click on a given website.
More informationSection 1: Introduction
Multitask Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design (1991) By Bengt Holmstrom and Paul Milgrom Presented by Group von Neumann Morgenstern Anita Chen, Salama Freed,
More information1.. There are two firms that produce a homogeneous product. Let p i
University of California, Davis Department of Economics Time: 3 hours Reading time: 20 minutes PRELIMINARY EXAMINATION FOR THE Ph.D. DEGREE Industrial Organization June 30, 2005 Answer four of the six
More informationRationing Poor Consumers to Reduce Prices
Rationing Poor Consumers to Reduce Prices Simona Grassi Ching-to Albert Ma Max Weber Fellow Department of Economics European University Institute Boston University Villa La Fonte, Via Delle Fontanelle,
More informationDo Ads Compete or Collaborate? Designing Click Models with Full Relationship Incorporated
Do s Compete or Collaborate? Designing Click Models with Full Relationship Incorporated Xin Xin School of Computer Science Beijing Institute of Technology xxin@bit.edu.cn Michael R. Lyu The Chinese University
More informationA Parametric Bootstrapping Approach to Forecast Intermittent Demand
Proceedings of the 2008 Industrial Engineering Research Conference J. Fowler and S. Mason, eds. A Parametric Bootstrapping Approach to Forecast Intermittent Demand Vijith Varghese, Manuel Rossetti Department
More informationSponsored Search Markets
COMP323 Introduction to Computational Game Theory Sponsored Search Markets Paul G. Spirakis Department of Computer Science University of Liverpool Paul G. Spirakis (U. Liverpool) Sponsored Search Markets
More informationA Survey of Sponsored Search Advertising in Large Commercial Search Engines. George Trimponias, CSE
A Survey of Sponsored Search Advertising in Large Commercial Search Engines George Trimponias, CSE Outline Introduction Structure of Sponsored Search Advertising (SSA) Practical Issues in SSA Historical
More informationAdWords Business Post-Campaign Summary Report - Fast Lane Clothing Co. Client Overview Campaign Overview: Fast Lane Clothing Company dba Tampa
AdWords Business Post-Campaign Summary Report - Fast Lane Clothing Co. Client Overview Campaign Overview: Fast Lane Clothing Company dba Tampa T-Shirts is located in Tampa, Florida. The company s primary
More informationMachine Learning Logistic Regression Hamid R. Rabiee Spring 2015
Machine Learning Logistic Regression Hamid R. Rabiee Spring 2015 http://ce.sharif.edu/courses/93-94/2/ce717-1 / Agenda Probabilistic Classification Introduction to Logistic regression Binary logistic regression
More informationOur main objective in this paper is to measure the value of customers acquired from Google search advertising
CELEBRATING 30 YEARS Vol. 30, No. 5, September October 2011, pp. 837 850 issn 0732-2399 eissn 1526-548X 11 3005 0837 http://dx.doi.org/10.1287/mksc.1110.0658 2011 INFORMS Measuring the Lifetime Value of
More informationOur main objective in this paper is to measure the value of customers acquired from Google search advertising
CELEBRATING 30 YEARS Vol. 30, No. 5, September October 2011, pp. 837 850 issn 0732-2399 eissn 1526-548X 11 3005 0837 http://dx.doi.org/10.1287/mksc.1110.0658 2011 INFORMS Measuring the Lifetime Value of
More informationThe Gender-Brand Effect of Key Phrases on User Clicks in Sponsored Search
The Gender-Brand Effect of Key Phrases on User Clicks in Sponsored Search Partha Mukherjee Pennsylvania State University. College of Information Science and Technology. University Park, PA-16802, USA pom5109@ist.psu.edu
More informationFOLLOWING THE CYBERSPACE BREADCRUMBS : MODELING OPTIONS AND INTERACTIONS AMONG CONSUMERS, ADVERTISERS AND SEARCH ENGINE PROVIDERS MING CHENG
FOLLOWING THE CYBERSPACE BREADCRUMBS : MODELING OPTIONS AND INTERACTIONS AMONG CONSUMERS, ADVERTISERS AND SEARCH ENGINE PROVIDERS by MING CHENG A Dissertation submitted to the Graduate School-Newark Rutgers,
More informationResponse Prediction in Performance based Display Advertising
Response Prediction in Performance based Display Advertising Krishna Prasad Chitrapura pkrishna@yahoo-inc.com Yahoo! Labs, Bangalore - 1 - Overview History of display advertisement From traditional print
More informationHarbingers of Failure: Online Appendix
Harbingers of Failure: Online Appendix Eric Anderson Northwestern University Kellogg School of Management Song Lin MIT Sloan School of Management Duncan Simester MIT Sloan School of Management Catherine
More informationBidding for Sponsored Link Advertisements at Internet
Bidding for Sponsored Link Advertisements at Internet Search Engines Benjamin Edelman Portions with Michael Ostrovsky and Michael Schwarz Industrial Organization Student Seminar September 2006 Project
More informationMKTG 555: Marketing Models
MKTG 555: Marketing Models Concluding Session April 25, 2017 1 Papers to Read Little (1970) Shocker and Srinivasan (1974) Guadagni and Little (1983) Roberts and Lattin (1991) Gupta (1988) Balasubramanian
More informationInequality and the Organization of Knowledge
Inequality and the Organization of Knowledge by Luis Garicano and Esteban Rossi-Hansberg Since the seminal work of Katz and Murphy (1992), the study of wage inequality has taken as its starting point a
More informationOnline advertising campaigns often consist of multiple ads, each with different creative content. We consider
Vol. 32, No. 5, September October 2013, pp. 753 767 ISSN 0732-2399 (print) ó ISSN 1526-548X (online) http://dx.doi.org/10.1287/mksc.2013.0802 2013 INFORMS Online Display Advertising: Modeling the Effects
More informationOnline shopping and platform design with ex ante registration requirements. Online Appendix
Online shopping and platform design with ex ante registration requirements Online Appendix June 7, 206 This supplementary appendix to the article Online shopping and platform design with ex ante registration
More informationThe 2x2 Exchange Economy. Joseph Tao-yi Wang 2013/10/2 (Lecture 8, Micro Theory I)
The 2x2 Exchange Economy Joseph Tao-yi Wang 2013/10/2 (Lecture 8, Micro Theory I) Road Map for Chapter 3 Pareto Efficiency Allocation (PEA) Cannot make one better off without hurting others Walrasian (Price-taking)
More informationOnline shopping and platform design with ex ante registration requirements
Online shopping and platform design with ex ante registration requirements O A Florian Morath Johannes Münster June 17, 2016 This supplementary appendix to the article Online shopping and platform design
More information1 Introduction 1. 2 Forecasting and Demand Modeling 5. 3 Deterministic Inventory Models Stochastic Inventory Models 63
CONTENTS IN BRIEF 1 Introduction 1 2 Forecasting and Demand Modeling 5 3 Deterministic Inventory Models 29 4 Stochastic Inventory Models 63 5 Multi Echelon Inventory Models 117 6 Dealing with Uncertainty
More informationEconomics Department 2009 Honors General Exam
Economics Department 2009 Honors General Exam This is a 3-hour exam. (For joint concentrators with economics as the secondary field, this is a 1-hour exam. Choose one section of the exam to complete, and
More informationHow Much Do We Know About Savings Attributable to a Program?
ABSTRACT How Much Do We Know About Savings Attributable to a Program? Stefanie Wayland, Opinion Dynamics, Oakland, CA Olivia Patterson, Opinion Dynamics, Oakland, CA Dr. Katherine Randazzo, Opinion Dynamics,
More informationThe Impact of SEM Programs on Customer Participation Dan Rubado, JP Batmale and Kati Harper, Energy Trust of Oregon
The Impact of SEM Programs on Customer Participation Dan Rubado, JP Batmale and Kati Harper, Energy Trust of Oregon Abstract Strategic Energy Management (SEM) is designed to positively change how customers
More informationMazzeo (RAND 2002) Seim (RAND 2006) Grieco (RAND 2014) Discrete Games. Jonathan Williams 1. 1 UNC - Chapel Hill
Discrete Games Jonathan Williams 1 1 UNC - Chapel Hill Mazzeo (RAND 2002) - Introduction As far back as Hotelling (1929), firms tradeoff intensity of competition against selection of product with strongest
More informationThe 2x2 Exchange Economy. Joseph Tao-yi Wang 2012/11/21 (Lecture 2, Micro Theory I)
The 2x2 Exchange Economy Joseph Tao-yi Wang 2012/11/21 (Lecture 2, Micro Theory I) Road Map for Chapter 3 Pareto Efficiency Cannot make one better off without hurting others Walrasian (Price-taking) Equilibrium
More informationMathematical Modeling of Competition in Sponsored Search Market
Mathematical Modeling of Competition in Sponsored Search Market arxiv:1006.1019v2 [cs.gt] 31 Aug 2010 Jian Liu Dah-Ming Chiu Department of Information Engineering The Chinese University of Hong Kong Email:
More informationSponsored Search: Theory and Practice
Sponsored Search: Theory and Practice Jan Pedersen Chief Scientist Yahoo! Search and Marketplace Group 1 10 August 2006 Outline Why online advertising is important A brief history of sponsored search An
More informationOnline Appendix Stuck in the Adoption Funnel: The Effect of Interruptions in the Adoption Process on Usage
Online Appendix Stuck in the Adoption Funnel: The Effect of Interruptions in the Adoption Process on Usage Anja Lambrecht London Business School alambrecht@london.edu Catherine Tucker Katja Seim University
More informationA Simple EOQ-like Solution to an Inventory System with Compound Poisson and Deterministic Demand
A Simple EOQ-like Solution to an Inventory System with Compound Poisson and Deterministic Demand Katy S. Azoury San Francisco State University, San Francisco, California, USA Julia Miyaoka* San Francisco
More informationMicroeconomic Theory -1- Introduction and maximization
Microeconomic Theory -- Introduction and maximization Introduction Maximization. Profit maximizing firm with monopoly power 6. General results on maximizing with two variables 3. Non-negativity constraints
More informationStructural versus Reduced Form
Econometric Analysis: Hausman and Leonard (2002) and Hosken et al (2011) Class 6 1 Structural versus Reduced Form Empirical papers can be broadly classified as: Structural: Empirical specification based
More informationOptimization of Click-Through Rate Prediction in the Yandex Search Engine
ISSN 5-155, Automatic Documentation and Mathematical Linguistics, 213, Vol. 47, No. 2, pp. 52 58. Allerton Press, Inc., 213. Original Russian Text K.E. Bauman, A.N. Kornetova, V.A. Topinskii, D.A. Khakimova,
More informationMixed Duopoly Under Vertical Differentiation
ANNALES D ÉCONOMIE ET DE STATISTIQUE. N 33 1994 Mixed Duopoly Under Vertical Differentiation Isabel GRILO* ABSTRACT. In this paper we study a vertically differentiated duopoly market with a profit maximizing
More informationEnergy Savings from Programmable Thermostats in the C&I Sector
Energy Savings from Programmable Thermostats in the C&I Sector Brian Eakin, Navigant Consulting Bill Provencher, Navigant Consulting & University of Wisconsin-Madison Julianne Meurice, Navigant Consulting
More informationLinkedIn and Google Adwords Practical Tips
LinkedIn and Google Adwords Daniella Boni Marketing Director Tom Mulligan Director, Product Management and RFP Callan Capital Brandes Investment Partners LinkedIn and Google Adwords Reaching Your Audience
More informationUniform and Targeted Advertising with Shoppers and. Asymmetric Loyal Market Shares
Uniform and Targeted dvertising with Shoppers and symmetric Loyal Market Shares Michael rnold, Chenguang Li and Lan Zhang October 9, 2012 Preliminary and Incomplete Keywords: informative advertising, targeted
More informationBid rent model for simultaneous determination of location and rent in land use microsimulations. Ricardo Hurtubia Michel Bierlaire
Bid rent model for simultaneous determination of location and rent in land use microsimulations Ricardo Hurtubia Michel Bierlaire STRC 2011 May 2011 STRC 2011 Bid rent model for simultaneous determination
More informationIntroduction to Analytics Tools Data Models Problem solving with analytics
Introduction to Analytics Tools Data Models Problem solving with analytics Analytics is the use of: data, information technology, statistical analysis, quantitative methods, and mathematical or computer-based
More informationCS425: Algorithms for Web Scale Data
CS425: Algorithms for Web Scale Data Most of the slides are from the Mining of Massive Datasets book. These slides have been modified for CS425. The original slides can be accessed at: www.mmds.org J.
More informationChapter 2. Practice Quiz -
Chapter 2 II. Demand A. Demand Shifters 1. Income 2. Prices of Related Goods 3. Advertising and Consumer Tastes 4. Popolation 5. Consumer Expectations 6. Other Factors B. The Demand Function C. Consumer
More informationECONOMICS AND ECONOMIC METHODS PRELIM EXAM Statistics and Econometrics May 2014
ECONOMICS AND ECONOMIC METHODS PRELIM EXAM Statistics and Econometrics May 2014 Instructions: Answer all five (5) questions. Point totals for each question are given in parentheses. The parts within each
More informationReducing Inventory by Simplifying Forecasting and Using Point of Sale Data
Executive Summary - Atul Agarwal and Gregory Holt MLOG 2005 Thesis 1 Reducing Inventory by Simplifying Forecasting and Using Point of Sale Data Introduction This thesis assesses the value to vendors of
More informationThe Implications of Alternative Biofuel Policies on Carbon Leakage
The Implications of Alternative Biofuel Policies on Carbon Leakage Dusan Drabik Graduate student Charles H. Dyson School of Applied Economics and Management, Cornell University, USA dd387@cornell.edu Harry
More informationMeasuring long-term effects in marketing P.M Cain
Measuring long-term effects in marketing P.M Cain Conventional marketing mix models are typically used to measure short-term marketing ROI and guide optimal budget allocation. However, this is only part
More informationHW3: AEM 4550 Economics of Advertising PART I ANSWERS
Objective: to analyze online advertising effectiveness. HW3: AEM 4550 Economics of Advertising PART I ANSWERS This question deals with the case Air France Internet Marketing: Optimizing Google, Yahoo!,
More informationACHIEVING MORE BY PAYING LESS? HOW RETAILERS CAN BENEFIT BY BIDDING LESS AGGRESSIVELY IN PAID SEARCH AUCTIONS
Association for Information Systems AIS Electronic Library (AISeL) Research-in-Progress Papers ECIS 2017 Proceedings Spring 6-10-2017 ACHIEVING MORE BY PAYING LESS? HOW RETAILERS CAN BENEFIT BY BIDDING
More informationTheory Appendix. 1 Model Setup
Theory Appendix In this appendix, we provide a stylized model based on our empirical setting to analyze the effect of competition on author behavior. The general idea is that in a market with imperfect
More informationExploitation and Exploration in a Performance based Contextual Advertising System
Exploitation and Exploration in a Performance based Contextual Advertising System Wei Li, Xuerui Wang, Ruofei Zhang, Ying Cui Jianchang Mao and Rong Jin Performance-based online advertising Performance
More informationA model of obfuscation with heterogeneous firms PRELIMINARY DRAFT
A model of obfuscation with heterogeneous firms Samir Mamadehussene October 13, 2015 PRELIMINARY DRAFT Abstract This paper analyses how the incentives for firms to obfuscate consumers vary with their brand
More informationVALUE OF SHARING DATA
VALUE OF SHARING DATA PATRICK HUMMEL* FEBRUARY 12, 2018 Abstract. This paper analyzes whether advertisers would be better off using data that would enable them to target users more accurately if the only
More informationDisentangling Prognostic and Predictive Biomarkers Through Mutual Information
Informatics for Health: Connected Citizen-Led Wellness and Population Health R. Randell et al. (Eds.) 2017 European Federation for Medical Informatics (EFMI) and IOS Press. This article is published online
More informationPrice of anarchy in auctions & the smoothness framework. Faidra Monachou Algorithmic Game Theory 2016 CoReLab, NTUA
Price of anarchy in auctions & the smoothness framework Faidra Monachou Algorithmic Game Theory 2016 CoReLab, NTUA Introduction: The price of anarchy in auctions COMPLETE INFORMATION GAMES Example: Chicken
More informationOnline Technical Appendix: Key Word. Aggregations. Adaptive questions. Automobile primary and secondary. Automobiles. markets
Online Technical Appendix: Key Word Aggregations Original Word Adaptive question design Adaptive questions Adoption process Advertising and media research Advertising response functions Advertising response
More informationCS269I: Incentives in Computer Science Lecture #16: Revenue-Maximizing Auctions
CS269I: Incentives in Computer Science Lecture #16: Revenue-Maximizing Auctions Tim Roughgarden November 16, 2016 1 Revenue Maximization and Bayesian Analysis Thus far, we ve focused on the design of auctions
More informationTargeted Advertising with Inventory Management
Targeted Advertising with Inventory Management David Maxwell Chickering Microsoft Research Redmond, WA 98052 dmax@microsoft.com David Heckerman Microsoft Research Redmond, WA 98052 heckerma@microsoft.com
More informationWe develop a structural demand model that endogenously captures the effect of out-of-stocks on customer
MANAGEMENT SCIENCE Vol. 56, No. 7, July 2010, pp. 1180 1197 issn 0025-1909 eissn 1526-5501 10 5607 1180 informs doi 10.1287/mnsc.1100.1170 2010 INFORMS Structural Estimation of the Effect of Out-of-Stocks
More informationPeriodic Measurement of Advertising Effectiveness Using Multiple-Test-Period Geo Experiments
Periodic Measurement of Advertising Effectiveness Using Multiple-Test-Period Geo Experiments Jon Vaver, Jim Koehler Google Inc Abstract 1 Introduction In a previous paper [6] we described the application
More informationSPRING 2017 AEM 4550: Economics of Advertising Prof. Jura Liaukonyte HW3
Ground rules for Homework 1-5: SPRING 2017 AEM 4550: Economics of Advertising Prof. Jura Liaukonyte HW3 1. HW is due at the beginning of class on the due date designated for each homework (see course website
More informationEcon 792. Labor Economics. Lecture 6
Econ 792 Labor Economics Lecture 6 1 "Although it is obvious that people acquire useful skills and knowledge, it is not obvious that these skills and knowledge are a form of capital, that this capital
More informationKey Results A website was established for the organization and was viewed by409 distinct visitors.
POST CAMPAIGN REPORT Executive Summary Campaign Overview The AdWords campaign for Jaffa Sea Scouts lasted 21 days (19/5-9/6) and cost $250.34. The major goal of the following campaign was to increase the
More informationWHICH ANALYTICAL METHOD BEST MEETS THE. By Robert Stratton Senior Group Director MarketShare: A Neustar Solution
CHALLENGE? WHICH ANALYTICAL METHOD BEST MEETS THE ATTRIBUTION By Robert Stratton Senior Group Director MarketShare: A Neustar Solution ? ADDRESSING THE ATTRIBUTION QUESTION Advertisers don t always have
More informationWeb Appendix. Oligopolistic Markets with Sequential Search and Asymmetric Information: Evidence from Retail Gasoline Markets
Web Appendix Oligopolistic Markets with Sequential Search and Asymmetric Information: Evidence from Retail Gasoline Markets Maarten Janssen Paul Pichler Simon Weidenholzer 28th July 2010 Department of
More informationChoice Modeling Analytics Benefits Of New Methods
Choice Modeling Analytics Benefits Of New Methods By John V. Colias Over the past two decades, the marketing research industry has witnessed rapid growth in the use of choice models. These types of models
More informationComputational Challenges for Real-Time Marketing with Large Datasets
Computational Challenges for Real-Time Marketing with Large Datasets Alan Montgomery Associate Professor Carnegie Mellon University Tepper School of Business e-mail: alan.montgomery@cmu.edu web: http://www.andrew.cmu.edu/user/alm3
More informationSessions 3/4: Summing Up and Brainstorming
Niehaus Center, Princeton University GEM, Sciences Po ARTNeT Capacity Building Workshop for Trade Research: Behind the Border Gravity Modeling Friday, December 19, 2008 Outline 1 The Gravity Model From
More informationWE consider the general ranking problem, where a computer
5140 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 54, NO. 11, NOVEMBER 2008 Statistical Analysis of Bayes Optimal Subset Ranking David Cossock and Tong Zhang Abstract The ranking problem has become increasingly
More informationOn Optimal Multidimensional Mechanism Design
On Optimal Multidimensional Mechanism Design YANG CAI, CONSTANTINOS DASKALAKIS and S. MATTHEW WEINBERG Massachusetts Institute of Technology We solve the optimal multi-dimensional mechanism design problem
More informationWhat Makes Them Click: Empirical Analysis of Consumer Demand for Search Advertising
What Makes Them Click: Empirical Analysis of Consumer Demand for Search Advertising By Przemyslaw Jeziorski and Ilya Segal American Economic Journal: Microeconomics, 2015 Miguel Alcobendas Yahoo! February
More informationMeasurement and Analysis of OSN Ad Auctions
Measurement and Analysis of OSN Ad Auctions Chloe Kliman-Silver Robert Bell Balachander Krishnamurthy Alan Mislove Northeastern University AT&T Labs Research Brown University Motivation Online advertising
More informationReverse-engineering AdWords Quality Score factors
Reverse-engineering AdWords Quality Score factors Google recently launched a new version of the AdWords API that allows users to download Quality Score information at scale, so contributor Brad Geddes
More information1.. There are two complementary goods that are consumed together (each individual good
University of California, Davis Department of Economics Time: 3 hours Reading time: 20 minutes PRELIMINARY EXAMINATION FOR THE Ph.D. DEGREE Industrial Organization June 22, 2004 Answer four of the six
More informationUNIVERSITY OF OSLO DEPARTMENT OF ECONOMICS
UNIVERSITY OF OSLO DEPARTMENT OF ECONOMICS Exam: ECON4137 Applied Micro Econometrics Date of exam: Thursday, May 31, 2018 Grades are given: June 15, 2018 Time for exam: 09.00 to 12.00 The problem set covers
More informationMATH 240: Problem Set 4
MATH 240: Problem Set 4 You may either complete this problem set alone or in partners. If you complete it in partners, only one copy needs to be submitted. Show all your work. This problem set is due on:
More informationComparing applied general equilibrium and econometric estimates of the effect of an environmental policy shock
Comparing applied general equilibrium and econometric estimates of the effect of an environmental policy shock Jared Carbone 1 Colorado School of Mines Frisch Centre for Economic Research University of
More informationOnline Model Evaluation in a Large-Scale Computational Advertising Platform
Online Model Evaluation in a Large-Scale Computational Advertising Platform Shahriar Shariat Turn Inc. Redwood City, CA Email: sshariat@turn.com Burkay Orten Turn Inc. Redwood City, CA Email: borten@turn.com
More informationGeneral Equilibrium for the Exchange Economy. Joseph Tao-yi Wang 2013/10/9 (Lecture 9, Micro Theory I)
General Equilibrium for the Exchange Economy Joseph Tao-yi Wang 2013/10/9 (Lecture 9, Micro Theory I) What s in between the lines? And God said, 2 Let there be light and there was light. (Genesis 1:3,
More informationAn Economic Framework for Information Platform
Globecom 2012 - Communications Software, Services and Multimedia Symposium An Economic Framework for Information Platform Yanjiao Chen, Qian Zhang Department of Computer Science and Engineering, Hong Kong
More informationTHREE LEVEL HIERARCHICAL BAYESIAN ESTIMATION IN CONJOINT PROCESS
Please cite this article as: Paweł Kopciuszewski, Three level hierarchical Bayesian estimation in conjoint process, Scientific Research of the Institute of Mathematics and Computer Science, 2006, Volume
More informationWhy Do Shoppers Use Cash? Evidence from Shopping Diary Data *
Why Do Shoppers Use Cash? Evidence from Shopping Diary Data * Naoki Wakamori Department of Economics University of Mannheim Angelika Welte Currency Department Bank of Canada The Second Conference on Economic
More information