Fun with Numbers Applied Economics in Retail BankingGar：數(shù)字的樂趣在零售bankinggar應(yīng)用經(jīng)濟(jì)學(xué)

1、Fun with Numbers: Applied Economics in Retail BankingGary W. Class1 February 20122Wells Fargo & Company is a diversified financial services company providing banking, insurance, investments, mortgage, and consumer and commercial finance through more than 9,000 stores and 12,000 ATMs and the Internet

2、 across North America and internationally.One in three households in America does business with Wells Fargo. Wells Fargo has $1.3 trillion in assets and more than 270,000 team members across our 80+ businesses.We first in market value of our stock among our U.S. peers as of December 31st, 2011.Our v

3、ision: “We want to satisfy all our customers financial needs and help them succeed financially.”Who is Wells Fargo?23Who is Gary Class & why should I listen to him?ACADEMIC:B.A. cum laude, University of Pennsylvania, 1982Major: English & EconomicsM.B.A. Haas School, University of California, Berkele

4、y, 1988Concentration: Finance PROFESSIONAL:Senior Vice PresidentInternet StrategyWells Fargo BankCOMMUNITY:Chair, Parks & Recreation Commission, City of Albany, California 344Who is Gary Class and why should I listen to him ? Developed at Wells Fargo:Branch & ATM Site Planning ModelsTeller Staffing

5、& Scheduling ToolsOnline Banking Customer Data WarehouseCustomer Behavior Predictive ModelsCustomer Behavior Across Multiple ChannelsAnalytics to make Anytime, Anywhere Banking Happen45Retail Financial Services involves families and small businesses. Organizations which can participate directly in s

6、ecurities markets (businesses, governments, institutions) are by definition excluded and are the province of “wholesale or investment banking.” Retail Banking Defined5Functions of Retail Financial ServicesPayments. The financial system must provide a mechanism for the transfer of money and payments

7、for goods and services. Risk Management. mechanisms to mitigate the financial risks faced by consumers, notably via insurance products. Borrowingadvancing funds from the future to today. The function of household credit encompasses short-term unsecured borrowing, longer-term unsecured borrowing and

8、secured borrowing.Saving / Investingadvancing funds from today until a later date. These products vary based on the intended time horizon, level and type of risk borne by the investor, tax treatment, and other factors.6Functions of Retail Financial Servicesare facilitated by Quantitative Models that

9、 leverage information technologyPayments. Extensive fraud risk prevention & mitigation models, most notably “Falcon” from FICO. Borrowingadvancing funds from the future to today. Centralized Credit Bureau data & Credit Risk Score-cards (e.g. the “FICO Score”).Saving / Investingadvancing funds from t

10、oday until a later date.Application of Modern Portfolio Management (mean-variance analysis and the Fama-French extension thereto). Risk Management. A critical aspect of the traditional deposit banking system is “delegated monitoring” where banks can gain incremental insight into a customers credit r

11、isk by carefully evaluating the customers usage of deposit accounts.7Applied Economics in PracticeIdentify a pressing Strategic Issue and where the issue is amenable to a systematic solutionB. Identify the appropriate analytical approach (i.e. Model) to address the IssueC. Specify, estimate & valida

12、te the Model D. Build Decision Support Tools based on the ModelE. Distribute the Decision Support Tools for use “in the field”Key External Resources to Leverage for (B ) & (C) Academia (D) & (E) Consulting Firms & Technology Vendors 8Applied Economics in Practice: Experimental DesignPROBLEM:Unlike i

13、n the natural sciences, there really are no repeatable, controlled experiments in the social sciences including economics. In the business world, customer behavior is influenced by a host of factors exogenous to the direct relationship of the firm with the customer: macro-economic factors, competiti

14、ve dynamics, the “social & cultural” calendar.SOLUTION:One best practice is to leverage “natural experiments” of policy changes or product introduction & use behavioral models as “controls” for confounding factors. Dennis Campbell Harvard UniversityJim Manzi, founder of Applied Predictive Technologi

15、es9Behavioral Model: Customer Attrition PROBLEM: Customers defect from the bank, taking their current & potential revenue stream with them.SOLUTION: Develop a model to identify the factors associated with retention and assess the risk that a customer will “attrite” (i.e. close all of their accounts

16、with the bank) in the forthcoming six months. APPLICATION:Tactical = Intervention & OutreachStrategic = Motivates the development of products & services that promote customer satisfaction & thereby maximize the “switching cost” for the customer to defect to another bank.10Customer Attrition Model :

17、Key FactorsRelationship with the Bank:What accounts do I have and how long have I had them?Product HoldingSet of products held with the bankBank Tenurelength of relationship Checking Account Activity:How actively am I using my checking account?Checking Account Balanceintervals, a proxy for “primary

18、bank relationship”Service Channel Behavior:How do I like to do my banking?Channel Activityidentifies willingness of customers to transact outside of branchesOnline Activity Segmentsbased on “functionality” and frequency of usageDemographics & Location:Who am I & where do I live?Customer TypeRetail o

19、r Mixed (i.e. owns business accounts, also)GeographyBased on customers residence, strength of Branch & ATM network11Logistic Regression Econometric ModelObjective: Score each customer based on the likelihood to discontinue banking relationship with Wells Fargo.Method: Perform univariate analysis and

20、 develop segmentation on selected customer attributes.Apply logistic procedure to estimated a binary choice model for whether or not “attrited” Where Pi is the attrition probability for customer i. is the intercept parameter. Zi is a vector of explanatory or independent variables for customer i.Attr

21、ited by 6 month?YesNo12Model Performance DiagnosticsThere is a well-defined methodology to assess the performance of logistic regression models. I. Classification TableOver-arching consideration is the ability of the model to predict the behavior of interest (in this case, customer attrition) in the

22、 population. The framework for evaluation is a “truth table” originally developed in the pharmaceutical domain. Accuracy is a key performance statistic - the higher the better. Obviously, no model is perfect. One goal is to balance the occurrence of “false positives” with that of “false negatives”.

23、This is addressed by the performance metrics of “sensitivity” and “specificity”; the key consideration is that the two values are balanced. 1313Model Performance Diagnostics I. Classification TablePredictedTrue PositiveFalse NegativeTrue NegativeFalse PositiveAttrition=1RetainedAttrition=0 (Retentio

24、n)ActualAccuracy: (true positives and negatives) / (total class)Error Rate: (false positives and negatives) / (total class)Sensitivity: (true positives) / (total actual positives) Specificity: (true negatives) / (total actual negatives)PredictedPositivePredictedNegativeActualPositiveActualNegativeAt

25、tritedAt a specific Probability Level Pcutoff , customers can be classified as Attrition =1 or Retention =1 and can be compared with the actual values whether the customer is attrited or retained.1414Model Performance DiagnosticsII. Dispersion (Gains Chart)A key performance consideration is the abil

26、ity of the model to discriminate the behavior of interest (in this case, customer attrition) and separate those likely to exhibit the behavior in the background population from those who are not. A popular way to visualize this is the Gains Chart which sorts the population into deciles and reports t

27、he ability of the model to identify the behavior of interest. III. Stability How well does the model deal with new sets of input data? Is the model stable over time? One method is to compare the Gains Chart of Forecast Validation dataset with the Gains Chart for the Estimation dataset. 1515Model Per

28、formance Diagnostics: Gains ChartCustomer Attrition, Predicted vs. ActualModel Score “Ventiles” grouped into Likelihood-to-Attrite SegmentsVentilesVERY LOWLOWMODERATEHIGHVERY HIGH16Applied Economics in PracticeCase Studies:Customer Satisfaction Measurement: Linkage of Behavioral & Attitudinal DataBr

29、anch & ATM Location Site SelectionOperations: Branch Teller Staffing & SchedulingDatabase Marketing: Harvesting the proliferation of internet data to improve leads for Branch Bankers17The overall goal is to identify the incremental ability of customer attitudes, as measured by market research survey

30、s, to enhance the models developed from Behavioral Data.First, we need to identify the salient question* in the customer satisfaction survey to use in the analysis.Next, we need to isolate the extent to which Customer Attitudes measured by the survey are actually related to future Customer Behaviors

31、*.Linkage of Behavioral Data & Attitudinal Data: Goals*via Factor Analysis, a statistical data reduction technique used to explain variability among observed random variables in terms of fewer unobserved random variables call “factors”. The observed variables are modeled as linear combinations of th

32、e factors, plus “error” terms. * 18Attitudinal Data from Customer Satisfaction SurveyA representative sample of Customers were asked, on a monthly basis, via email: Q1: How satisfied are you with Wells Fargo?Five point scale, where 1=Not Satisfied and 5=Extremely Satisfied19Actual Customer Retention

33、 and self-reported Customer Satisfaction are higher where the Predicted Risk of Customer Attrition is lowerPredicted Risk of Customer AttritionVery High High Moderate Low Very Low3.784.134.104.234.37 Actual Customer Retention %Mean Response to Q1: How Satisfied are you with Wells Fargo?Linkage of Be

34、havioral Data & Attitudinal Data: Findings20Applied Economics in PracticeCase Studies:Customer Satisfaction Measurement: Linkage of Behavioral & Attitudinal DataBranch & ATM Location Site SelectionOperations: Branch Teller Staffing & SchedulingDatabase Marketing: Harvesting the proliferation of inte

35、rnet data to improve leads for Branch Bankers2122Wells Fargo 4th & Brannan Branch in San Francisco is co-located with a Starbucks22Branch & ATM Location Site SelectionPROBLEM: What is the optimal distribution network (branches, ATMs, etc.) to cultivate existing customers & acquire new ones? SOLUTION

36、:Financial ModelingEconomics of the discreet bank branch location, focused on customer “patronage” Strategic MarketingWhat markets to serve, what street-corners to be on & what customers will visit the location?Real EstateWhat is the marketplace value of this unique location?Avijit Ghosh University

37、of Illinois23PROBLEM: How can we dimension household banking behavioral “spatially”SOLUTION: Develop & describe customer branch & ATM data as a “visitation matrix”. This allowed assignment of customers to individual branch locations via “patronage” and the delineation of “empirical trade areas” for

38、individual branches allowing a precise estimation of local product demand.Branch & ATM Location Site Selection24Customer Behavior & Branch Empirical Trade Area25ATM Locations: How do they provide value to Wells Fargo? the fee revenue that WFC earns when other banks customers pay to use WFC ATMs.the

39、incremental customer satisfaction stemming from convenient cash access points.26“Clump Analysis” Provides a Visual Indication of Unmet demand for ATM Withdrawals Non-WFC ATM transactions illustrate concentrations of WF customer activity and commerce.ATM Locations: Identification of Demand27Applied E

40、conomics in PracticeCase Studies:Customer Satisfaction Measurement: Linkage of Behavioral & Attitudinal DataBranch & ATM Location Site SelectionOperations: Branch Teller Staffing & SchedulingDatabase Marketing: Harvesting the proliferation of internet data to improve leads for Branch Bankers28Operat

41、ions: Branch Teller Staffing & SchedulingPROBLEM: Mis-match between the availability of tellers and customer demand for teller services.This is a “dead-weight loss” as idle tellers waste the banks resources and waiting customers waste customers patience.Direct Cost of Teller TimeIndirect Benefit of

42、Customer Retention via Quality Customer Service29Operations: Branch Teller Staffing & SchedulingSOLUTION:Forecast the customer demand for teller services for each branch each day for each half-hour using historical data.Leverage queue-ing models to determine the level of teller staffing required to

43、meet the forecast demand, with acceptable customer wait times.Develop software for branch managers to “drag & drop” branch employees on their roster into a schedule sufficient to meet the forecast demand and minimize customer wait time.Ali Kiran founder of KCG Consulting 30Staffing & Scheduling Mode

44、l ComponentsForecasting: start with forecasting the three building blocks:* arrival rates, * service time and * customers (im)patience, combining the first two to create forecast of the offered-load (or workload). Staffing: Identify the costs constraints acceptable level of service; then apply queue

45、-ing models, subject to these constraints, in order to design half-hourly staffing levels. Shifts: Integer programs, or combinatorial optimization, is then used to aggregate hourly demand into shifts. The output is a shift-schedule, specifying how many agents should available to occupy each shift Ro

46、stering: Out of the body of workers, who should come to work and when? Source: Avi Mandelbaum, “Service Engineering: Data-Based Science & Teaching in support of Service Management. Technion, Haifa, Israel, 2006.31Queue-ing Models: Erlang-AThe most widely adopted methodology for staffing models is th

47、e algorithm developed by Erlang, which require the identification of four key parameters.Source: Avi Mandelbaum, “Service Engineering: Data-Based Science & Teaching in support of Service Management. Technion, Haifa, Israel, 2006.32To dimension “abandonment”, an R&D effort included experimentation wi

48、th software algorithm that converts video to anonymous branch visitor tracking data in real-timeRalph Crabtree & Iain Currie, founders of Brickstream Branch Teller Staffing: Impatience & Abandonment33Applied Economics in PracticeCase Studies:Customer Satisfaction Measurement: Linkage of Behavioral &

49、 Attitudinal DataBranch & ATM Location Site SelectionOperations: Branch Teller Staffing & SchedulingDatabase Marketing: Harvesting the proliferation of internet data to improve leads for Branch Bankers34Database Marketing: Harvesting the proliferation of internet data to improve leads for Branch Ban

50、kersBACKGROUND:In the 1990s, the focus was on leveraging customer profile and account activity to develop models to predict customers likelihood to purchase another product and responsiveness to direct marketing. PROBLEM:The deployment of internet banking applications generated a humongous amount of

51、 customer interaction data - how could we make sense of it & use it to identify leads for branch bankers?SOLUTION:Leverage Data warehousing technology to join Customer Profiles, Model scores with granular “secure banking application” & “public site” navigation.Dirk van den Poel Ghent UniversitySteve

52、 Krause, founder of Personify3536Take the customers web browsing & online banking activityPeter Heffring, founder of Ceres Marketing Systemsfind the customers whose behavior indicates a sales lead and distribute the leads to bankers in the branch36“Drowning in Data”PROBLEM: The emergence of the inte

53、rnet has contributed to a proliferation of very complex data ripe for analysis* Structured Data (Administrative & Accounting)* Semi-Structured Data (system logs & weblogs)* Unstructured Data (Text, Speech, Image)37“Drowning in Data”, a partial solutionIn order for the bank to leverage & act upon sem

54、i- or unstructured data, it is convenient to transform it into relational data by: Customer-ization = explicitly linking the activity to a known customer Session-ization = creating logical groupings of the stream of interactions by user_agent & time_stampNote: a special case of session-ization are b

55、anker-to-customer interactions where the data is modeled to approximate the customers view of the interaction, i.e. a “sojourn”.This approach yields cross-sectional, longitudinal time-series data by unique customer opening the door to the next generation of quantitative methods in applied economics,

56、 notably Behavioral Economics.38Behavioral Economics integrates economics and psychologyThe discipline focuses on:1. Identifying the ways in which behavior differs from the traditional model of economics2. Showing how this behavior matters in economic contextsIt seeks to explain why people dont alwa

57、ys make rational decisions Whats Hot: The emergence of Behavioral Economics39What is Behavioral Finance?Leverages Behavioral Economicsthe introduction of Psychological methods in Laboratory Experiments of Decision Making and increased awareness of “cognitive bias” that challenges the “rational decis

58、ion maker” of neo-classical economicsDaniel Kahneman and Amos Tversky It is a branch of Household Finance EconomicsHow do individuals / families make financial decisions as compared to Institutions?A key insight that Household Finance decisions are innately more complex than those in Corporate FinanceA great overview was presented in John Y. Campbells presidential address to the American Finance Association on January 7, 2006.40Whats Hot in Marketing Science: PROBLEM: In “contractual” settings