Data-Driven Decisions: Role of Operations Research in Business Analytics Dr. Radhika Kulkarni Vice President, Advanced Analytics R&D SAS Institute April 11, 2011
Welcome to the World of Analytics! Lessons learned in the journey from OR to Business Analytics What is Business Analytics? Why should OR professionals care? Why should businesses care? Responding to the market trends Educating the next generation How can we ensure a prominent place for OR in Business Analytics? 2
Lessons learned in the journey from OR to Analytics Why should the world care about OR? Where can OR add value? What can we do more of? What more can we do? Value of Optimization Solutions versus OR tools Integrating into the bigger picture Being essential to the big picture 3
Organizations Have Lots of Data, Data, Data ERP Systems Other Operational Apps Web Logs etc Other Legacy Operational Systems Apps Call Centre Apps Operational Switches File based information Unstructured Data Operational Switches 4
Information Explosion 5
Harnessing Value from the Data Data Deluge Data-Driven Insights 6
Global Challenges Proactive Decision Making using Advanced Analytics What happened? Where exactly is the problem? What if these trends continue? How many, how often, where? What s the best that can happen? What will happen next? What actions are needed? Why is this happening? 7
Business Analytics What is it? The use of statistical analysis, data mining, forecasting, and optimization to make critical business decisions based on customer and operational data. Why should we care? Increased recognition of the value of OR tools as an integral part of a comprehensive solution capitalize on it! There is tremendous opportunity to add value in multiple disciplines grab it! Business applications of OR can reach a wider audience exploit it! It is an opportune time for ensuring OR is recognized as a key component of Business Analytics 8
The Whole is Greater than the Sum of its Parts Each technology works well on its own, but combining them all is the real opportunity Forecasting Quality Improvement Service Intelligence Statistical Modeling Optimization Forecasting Service Parts Optimization Optimization Data Mining Statistical Modeling Quality Improvement Data Management Data Management 9
Example: Markdown Optimization in Retail Department stores and fashion retailers clear apparel inventory by reducing prices What prices should we set through the end of the season to meet inventory goals and maximize revenue? Typical Retail Merchandise Hierarchy: Department Distributed data Large volumes of data Algorithmic complexity Class Subclass Style 10 10
Business Problem Determine a price schedule in order to liquidate inventory profitably by a certain date for a variety of business reasons Drop prices increase demand (hopefully) When should I give a discount? How much? What happens to my margin? Initial Inventory week 1 week 2... week T p 1 p 2... p T Target Inventory D 1 D 2... D T 11
Why an Analytical Solution? Huge number of decisions 50,000 products/store x 2,000 stores = 100,000,000 decisions! Large volume of weekly data for two-year moving window Number of units sold, price at point of sale, reference price Starting and ending inventory in each store Promotions, TV ads, circulars, Building blocks Statistical model for demand Price optimization 12
Trends: Distributed Computing Gridded environments and multicore processors are increasingly cost effective Performance is gained by breaking work into tasks that can be done in parallel by nodes or processes 13
Markdown Optimization High Performance Computing Customer scenario: More than twelve billion dollars in markdowns last year 273 million product-by-location combinations Hundreds of millions of pricing decisions per week Three terabytes of historical sales data. Cut markdown optimization compute time from 31 minutes to 3.5 minutes Reduction in hardware costs by 80% Speed of computation enables multiple scenarios to be run potential to change the game! 14
Markdown Optimization Lessons Learned Value in being part of an integrated solution Optimization technique and estimation functions need to work together The analytical pieces are not the computational bottleneck: Data I/O is. Reminds you that there is a great deal of uncertainty in the parameters that go into the optimization problem: Is it that important to solve to the nth degree of optimality? There are continued challenges which spark innovation that can provide game changing solutions 15
Distributed computing Where else will we see need? Risk, medical applications, social network applications, text search, retail, financial applications,. What are the challenges with distributed computing? Data storage Different algorithms Return to look at dense techniques for some problems? Preference for solution techniques that exploit data structures 16
Pushing the envelope in multiple disciplines Forecast reconciliation elegant application of QP Typical Retail Merchandise Hierarchy: Department Class Subclass Style 17
Pushing the Envelope in Multiple Disciplines: Optimal Binning in Credit Scoring Optimal binning for Credit Scoring: Group credit applicant risk based on attributes Based on good / bad trends demonstrated by data Variables: AGE, INCOME, etc. Aggregation by variable ranges Example: AGE groups [20-29], [30-39], [40-49], [50-59] Accuracy of risk measures heavily depend on binning What is the best binning choice?» The one that minimizes deviation from Weight of Evidence (a measure of risk) Nonlinear regression with binary decisions 18
Optimal Binning in Credit Scoring Example: AGE attribute Constraints:» At most 7 bins» Max bin widths» Min bin widths» Monotonic increasing Minimize deviation from WOE data (green) 19
18 of 31 Marketing Optimization - Problem Description Cost=$3.00 Expected Return=$5.50 Cost=$2.25 Expected Return=$4.90 Cost=$1.00 Expected Return=$3.90 Visa Classic / Direct Marketing Visa Classic / Call Center Visa Classic / Branch Visa Gold / Direct Marketing Visa Gold / Call Center Visa Gold / Branch Visa Value / Direct Marketing Visa Value / Call Center Visa Value / Branch Customers Products x Channels Common problem in multiple applications 20
Cross-Sell / Upsell opportunity Integrate optimization into campaign management Business Issue Grow market share Increase revenue per acquired customer Retain and upgrade existing customers Results/Benefits 30% reduction in campaign costs 3-10x better response rates 4x Campaign ROI 3 month implementation ROI in months 21
Call Center Offer Optimization Expanding the reach of analytical solutions Provider of Digital TV, broadband Business Issue Competitive market Use every contact point to cross-sell / upsell Solution Deliver analytical results to call center employees Results/Benefits Increased customer loyalty and retention Increased revenue Democratization of Analytics? 22
Operationalizing Analytics Scalability Business Issue Analyze massive amounts of data to accurately predict which customers would be most likely to respond to targeted marketing offers. 250M transactions / week Goal Deliver the right offer to the right customer, right at the point of sale Meeting performance requirement Exploit in database analytics 23
Key Takeaways from the Last 3 Examples Gain value by being part of an integrated solution Optimization parameters obtained as scores from a data mining model Results of optimization feed back into a campaign management system Expand reach by operationalizing the application Present offers at the right time via the call center Customized offers presented at the checkout counter Deliver the right offer to the right customer, right at the point of sale. Exploit in-database computations to minimize data movement (for the input scores to the optimization model) 24
Trends: Operationalizing analytics Where will we see need? Call center offers Real-time offers in retail Credit card fraud prevention Instrument solutions to self-monitor What are the challenges? Ability to embed algorithms in a business process Delivery mechanism for Software as a Service Automation of a workflow with ability to run instantaneous analytical decision support analytical engines 25
Trends: Integrating Modeling & Visualization Social Network Analysis Applications Fraud detection Terrorist network detection Targeted marketing Software needs Visualization Grid computing Algorithms 26
Trends: Use of Unstructured Data Unstructured data Structured data 25% 5% 70% Semistructured data 27
Uses for Text Analytics Text Analytics has numerous applications Retail Identify the most profitable customers and the underlying reasons for their loyalty. Finance Retention of current customer base using call center transcriptions or transcribed audio. Insurance Identify fraudulent claims. Track competitive intelligence. Brand management Manufacturing Reduce time to detect root cause of product issues. Identify trends in market segments. Government Early warning of terrorist behavior Healthcare fraud Corruption Life Sciences Identify adverse events. Recommend appropriate research materials. 28
What Does All of This Mean for Us? 29
Data Are Driving the Demand for Analytical Experts Critical business problems are increasingly characterized by large volumes of data and the need for rapid analysis. Organizations are competing on analytics, and they value data and analytics as strategic assets. There is a global demand for analytical experts who can Formulate and solve problems using a broad ensemble of methods Collaborate with interdisciplinary teams on analytical software solutions Davenport and Harris 30
What Are the Challenges for OR Practitioners? Learn to work with new sources and new types of data Transactional Customer Textual Become more involved in making the data analysis-ready Exploit new analytic infrastructure for large data Distributed In-database Collaborate with interdisciplinary teams that combine Industry experience Specialized software development skills for data management, analytical computing, reporting Expertise in statistics, data mining, forecasting, optimization 31
Remember What Customers Need Effective Visualization Deployment More Powerful Algorithms More Flexible Models 32 32
Educate the next generation Need to get it right OR programs should be (real-world) business-focused INFORMS can set this tone at the academic level Embrace all areas of analytics data integration, forecasting, predictive; not just optimization and simulation. Equip students to formulate and solve problems with large and diverse data sets Train more students in cross-discipline skills and computational techniques including knowledge of machine architecture and how to exploit it Partner with business and industry Internships Funding for new programs and research, including applied work 33
What Does All of This Mean for Us? The opportunities for Operations Research practitioners are unprecedented! We need to ensure that OR earns it rightful place as a key discipline in Analytics 34
Radhika.Kulkarni@sas.com