Abstracts

Survival analysis, also called time-to-event analysis, is an underutilized tool in the data miner's toolkit. Because many analytical practitioners in the business world are unfamiliar with survival analysis, they have become adept at reformulating questions that should properly be about "when" into questions about "whether."

A typical attrition model asks "Will the customer still be active in six months?" This is a convenient question because it has the kind of yes or no answer that can easily be modelled using logistic regression, or decision trees, but a much more interesting question is "How long will the customer last?" The answer to that question has many uses from lifetime value calculations to optimizing acquisition channels.

The particular application addressed in this talk is creating a long-range daily forecast for a subscriber population that is a constantly changing mix of customers segments, each with its own survival function. A bottom-up forecast of this sort is more useful as a planning tool than the typical aggregate forecast because the survival-based forecast is very sensitive to changes in assumptions about the characteristics of new customers.

The analyst can study the effect of varying assumptions about credit scores, product preferences, rate plans, demographics, or whatever else is of interest by watching the effect of these assumptions on the forecast. The work described here is based on several successful forecasts developed for newspapers, telephone companies, and other subscription-based businesses. Most company forecasters regularly use their judgment to adjust the statistical forecasts that are generated by their software. But when are such interventions effective and when are they likely to damage accuracy?

Paul Goodwin will in this presentation use an analysis of over 60,000 company forecasts to identify when judgmental adjustment is justified and when it is not. It will show how forecasters can make better use of their time by confining their attention to occasions when important events will have an impact on statistical forecast accuracy and by avoiding frequent small adjustments.

He will also look at the biases that are associated with judgmental adjustments to forecasts, such as optimism bias and recency bias and will discuss how these biases can be avoided. We need data in order to make the best possible decision, and this requires us not just to have all the figures but presenting them so they make an impact. Lomborg will look at the issue of Global warming to show not only how it is real, but also how we are often presented with poor, one-sided and exaggerated analysis. This leads to bad policies that will do little to tackle the warming at extremely high cost, as in the Kyoto protocol and the new EU promises. We need smarter solutions focused on getting long-term impacts through increased funding of R&D for zero-carbon energy. And finally, we should remember, that if we really want to help the world, there are many other and better things we could focus on first, like malnutrition, free trade, vaccines, agricultural technology, education etc. After a long history of subjective, bottom up forecasting on many fronts, purchasing, supply chain, commercial, and finance, Dow is now progressing to statistical forecasting. Though univariate forecasts are in fact useful in many situations, longer term multivariate forecasts generally have more utility for strategic unconstrained forecasting. Implementing large scale statistical forecasting is not just a software play. Large numbers of potential predictor variables are generally proposed by business experts and thus advanced data mining techniques are necessary to pre-process the data to reduce the size of the problem. People, processes and technology, in the form of data and software, all have to interact together to be successful. Infrastructure details and examples of successful large scale forecasting projects will be shared.

Glass production was known to the Romans, and the manufacturing of sheet glass is about 200 years old. This presentation shows how a data-driven approach coupled with visual analysis in JMP can improve even this mature, well-understood process. Using live examples, you will gain an insight into how Saint Gobain has been able to make significant cost savings by reducing energy consumption whilst maintaining glass quality. SAS' prevalence within pharmaceutical companies has historically sat in the Clinical Trials area of the business; however trials are not the only division that can benefit from applying analysis to enhance critical business decisions. They have a wealth of experience in analysing controlled, accurate and complete data. However, applying their understanding of analysis to key insight such as physicians are most likely to prescribe is conversely very under utilised in the Sales and Marketing function of the business. With the use of powerful Data Mining tools, pharmaceutical companies can access this insight and leverage analytical techniques for business operational decision making.

The purpose of his paper is to talk through a particular project that took place at Eli Lilly, one of the world's largest research based pharmaceutical companies. The objective of the analysis was to understand causes of considerable variability in sales across the UK for a particular product. The focus being given to the modelling approach taken, data challenges faced and lessons learned; one key challenge being the definition of a reliable and unbiased target variable. The German social insurance system underwent a major reform in 2009. A central health fund has been installed to collect payments from the insured, employers and government subsidies. Sickness funds will receive risk-adjusted payments from the central health fund for enrolled members. The risk adjustment is based on a model including 80 chronic, severe and costly diseases. The model was developed from a sample of approximately 5 million members of the German sickness funds and will be applied to all 70 million insurants.

The presentation will cover the following points:

• Overview of the German social health insurance system and recent reforms.
• escription of how 80 diseases were chosen.
• Incorporation of diagnoses and drug prescription data.
• Calculation of risk-based payments to sickness funds.
• Further development.

Axel Springer is Germany's largest newspaper and third-largest magazine publisher, as well as one of the leading European media enterprises. With over 170 newspapers and magazines, more than 60 online offerings for various interest groups and information needs, as well as its holdings in television and radio stations, Axel Springer is active in a total of 35 countries. The flagship product is BILD, the largest newspaper in Europe.

The newspaper is printed at eight main printing plants and is distributed via 73 wholesalers to 120,000 retailers every night. Along this distribution chain, the publisher has to determine the right amount of copies for each wholesaler and the wholesaler has to deliver the right amount for each retailer. That means there will be lost earnings if customer demand is higher than the delivery amount. Likewise, the costs for returned copies will increase if the demand is less than the delivery amount.

Axel Springer has developed statistical methods for both steps of the supply chain to predict an optimised amount of copies.

The system VERITA assists the publisher's staff in evaluating the right amount of copies for the wholesaler. Trends, seasonal ups and downs, effects from special events and lost-sales estimation are KPIs of this process. VERITA helps to get more earnings by producing a better distribution across the wholesalers in a shorter time.

Predicting newspaper sales rates on the retailer level is another challenging task since the set of retailers is very heterogeneous. The regulation system RAMBOS uses neural nets to predict the future sales rates. Based on these forecasts, RAMBOS applies a rule-based approach generated by data mining techniques to optimise the individual supply for each retailer.

RAMBOS reduces the number of returns significantly without increasing the number of sell-outs. With the advent of the Internet, credit card companies began facing new challenges in fraud prevention. While our customers may shop anywhere, anytime, so may fraudsters. They are highly connected, well-organized and technically skilled. They focus on stealing magnetic stripe data at the point of sale or bulks of data from successful data hacks. The data is then sold through the Internet.

Early detection of fraud cases can save credit card companies significant amounts of money: A dollar not fraudulently spent is a dollar earned. Contrast this "profit margin" to the 1 percent to 2 percent margin of honest credit card usage.

At Viseca we employ a wide arsenal of techniques to detect fraud. One of these techniques is data mining (predictive modeling). Our biggest technical challenge comes from the fact that, luckily, fraud is still a rare event.

In this presentation Marcel Bieler will report on the company's quantitative methodology, the achieved precision, and its financial impact on earnings. In a telecommunications network, the improvement in the capacity planning process (planning, analysis and design activities essential in order to meet the demand for on-time services at a reasonable cost for the TLC operator while maintaining suitable quality levels for the customers) is becoming more and more important.

As the traffic volume grows, it is necessary to allocate in advance the right number of resources in order to prevent deterioration in the quality of service while keeping costs at a reasonable level.

The typical capacity planning process in a TLC network starts at the analysis of the network traffic, proceeds with the calculation of performance indicators and continues with the development of traffic forecasts on the provided services. These forecasts are then used to estimate the future impact of different volumes of services on the network.

Telecom Italia Network uses a trial SAS Forecast Server to verify the improvement of forecasting quality, using tools to support and automate some tasks.

Dino Cannas will show how and where (in different kinds of TLC networks) SAS Forecast Server can improve the quality of forecasting. British Airways has been using data mining techniques for over twenty years and Enterprise Miner for ten years. Descriptive and predictive data mining techniques have been used in a variety of applications in the business, principally in customer relationship management and in analysis of engine data, but also analysing passenger arrival patterns and website visit data and in understanding fraud risk.

The presentation will give an overview of data mining at British Airways and illustrate the variety of its applications. A case study will be presented describing the use of cluster analysis for re-designing the structure of intranet pages for employee information at British Airways. Focus groups were held where employees subjectively clustered topics and headings into categories. Cluster analysis in SAS was then used to create a tree structure on which to base the re-design of intranet pages. This made the information more accessible and more logically organised. Often in data mining problems we have large data sets with many variables and many missing data. The correct approach to handle this situation depends on the kind of data and the aim of the analysis. However we simply cannot delete incomplete records because this amounts to a substantial loss of costly collected data. Suppose for example that our data include 10 variables and that each of these has randomly 10% missing values. Then, on average 65% of the units will have at least one missing value.

We compare several approaches, which can be easily implemented in SAS. Single imputations, multiple imputations and non-parametric methods will be considered with an application to a European statistical survey. By knowing in advance the changes in the environment in which our businesses operate, we are able to act earlier and utilize the impact of a boom or a slowdown faster. In the last couple of years Danfoss has used prediction as a fact-based second opinion to our demand planning and management view on the future of the business. In the past the prediction was mainly due to the decision maker's intuition, emotions and personal experiences. Today a combination of the two are at work and has proved more effective both operationally and strategically.

Prediction has improved the outcome of our demand planning and has also reinforced our managerial decisions. It challenges the detailed planning of the existing order intake and provides product-related input, while formalizing the intuitive skills of managers and provoking their mindsets.

Danfoss' prediction model is based on external indicators in combination with turnover data from the business. The prediction uses a variety of statistical models that rely on historical data and/or causal variables that are drivers of the business.

The main focus of this presentation will be on how prediction is used at Danfoss. The basic idea of the Prestiamoci.it platform is to combine the innovative power of the Internet with an ancient and traditional kind of transaction: person-to-person lending.

The role of Prestiamoci.it is to enable and promote these transactions, creating a community and marketplace where two different and important needs meet: on one side to get a loan for personal or entrepreneurial purposes at an affordable rate, on the other side to make an informed investment. These needs are aligned with the lender's areas of interest, reducing the risk through many-to-many relationships.

Using SAS operational research tools, we have implemented marketplace optimisation processes that on a flexible, daily basis seek the best match between loan requests and money offers. As a result, multiple objectives - efficiency, timeliness and costs ? are satisfied for both the members and platform. Moreover, many constraints are enforced in order to satisfy platform policies and member requirements to invest in projects with strong affinity. Germany's telecommunications industry is facing a situation of increasing competition in a saturated market. This leads to a high level of pressure on prices, and therefore directly affects revenue and margins. Accurate planning, as well as early identification and determination of prospective developments, trends and changes, are essential in order to remain competitive and not to stay behind.

Using analytical methods, the BI department of Telefónica O₂ Germany supports business units with reliable predictions regarding the relevant KPIs of the company.

The intention of this presentation is to show the potential, as well as the business value of analytical predictions, and to describe the way they can be used to reduce the uncertainty of budget planning and tracking. During his presentation, Andreas Gimber will reflect the evolution of the planning process, from solely using descriptive historical data up to highly sophisticated forecasting using SAS Forecast Server. Furthermore, he will explain how to deal with upcoming challenges and illustrate an approach for achieving more efficient decisions by combining analytical knowledge and expert know how. The predictive power of product-affinity and churn-risk scores is crucial for the success of marketing campaigns. This is even more true when more technical approaches use the scores as input. Examples are next-best-offer algorithms and tools such as SAS Marketing Optimization, which rely heavily on the probabilities generated by scoring. How can the predictive power of the scores be measured and monitored? Is there a way to generate detailed, score-specific hints for improvements from this validation?

This talk presents the current state of the solution to these problems at Deutsche Postbank AG and includes some remarks on the efforts to score-back test of the next-best-offer algorithm. Hunziker will describe the project that enables the Luzern Canton Hospital (Switzerland) to deal automatically with medical knowledge from the beginning of a patient's treatment by making the data available in a structured way so that ICT systems can handle it as needed. With this view, the hospital has developed a solution together with SAS that "reads" doctors' reports and translates them into the necessary software codes for the system to operate as described. This model has been enhanced to the entry report in order to generate a complete patient treatment picture describing and structuring the patient's path through the duration of hospitalization.

Today, much of a doctor's knowledge about the treatment path of a patient is covered in medical reports, such as documents describing the diagnosis and procedures followed. Only a few clinical information systems today can handle parts of this knowledge as parameters, which are manually added to them. Additionally, in most hospitals DRG coding is a process regarded necessary only for reimbursement purposes. This is why most hospitals do coding only on the basis of doctors' exit reports (a manual work). However in this way, neither the hospitalisation process nor the quality and risk analyses of the patient's treatment can really rely on the knowledge contained in the doctors' reports. Enhancing operation systems, such as clinic information systems, with this kind of knowledge requires that systems be developed to be able to handle this information. At the beginning of the presentation Kosinski and Stepien will share their approach to customer retention. This approach embraces gathering business knowledge of the churn phenomenon, deriving reasons and preceding events of customer churn from a business perspective, and then designing data transformations and predictive variables with regards to data availability and quality. They also leverage SAS and Teradata in customer retention solutions. A focus will be placed on how SAS with Teradata allows organizations to extract information from a plethora of available raw traffic, CRM and financial data to forge data transformation, providing powerful predictive variables. They will present case studies of successful data transformations. Finally, Kosinski and Stepien will demonstrate how the agility of the SAS environment allows them to develop an entirely dynamic architecture of their application. The solution enables advanced business users to comprehensively manage the scope of performed data manipulation – expand, change and add new data aggregations and transformations without any intervention in application code. Hear the benefits of the SAS environment ? its flexibility and agility allowing business-oriented, dynamic-scope data mining applications with exceptional predictive power. TeliaSonera is the leading Nordic and Baltic telecommunications operator. After several years of good experiences on predictive modeling, the Swedish Broadband marketing unit decided to further improve direct marketing by using design and analysis of experiments when selecting target customers. This presentation focuses on sharing experiences from development work until present. The presentation includes a concrete case. Tools used include SAS QC/ADX, SAS Enterprise Miner, SAS Campaign Studio, and SAS STAT. ASL of Pavia is a public company, and as part of the wider regional healthcare system of Regione Lombardia, is responsible for healthcare governance in the province of Pavia.

According to its mission, ASL of Pavia has to guarantee that high standards of appropriateness, efficiency and efficacy are constantly pursued by service providers in order to meet citizens' needs. Therefore, ASL has to anticipate epidemiological trends in order to better focus strategic actions and address the available resources.

Based on an original data warehouse project, developed on SAS technology, ASL of Pavia has recently begun a SAS Forecast Server implementation to improve prediction capability in relevant informative areas.

Preliminary analysis, which focused on resource planning, showed that models based on SAS Forecast Server may significantly improve prediction capability compared to traditional ones; therefore, it is necessary to make an accurate evaluation of every single node in which the forecasting model is articulated.

Chronic disease management is a crucial target of healthcare and governing systems. ASL of Pavia is addressing it with a number of strategic actions. Accordingly, SAS Forecast Server is expected to support the analysis and prediction of relevant variables to drive their implementation.

Preliminary results were discussed during the Italian SAS Forum held in Milan in October 2008. Each day, nearly 30,000 aircrafts take to the skies of Europe moving 2 million people and tonnes of high-value goods. Achieving a safe, orderly and efficient flow of traffic while minimizing the impact on the environment requires a complex, layered approach to management, and planning on time scales from seconds ahead to decades. The job of EUROCONTROL is to coordinate air traffic management across the skies of its 38 member states. One of the ways in which this is accomplished is to provide forecasts of future traffic.

For more than 30 years EUROCONTROL has been providing annual flight forecasts, but in recent years the demands have grown for forecasting more frequently and at ever-finer resolution: by month, week, state, flow, airport, aircraft type or market segment.

EUROCONTROL faces challenges many forecasters will recognize: how to manage the flows of data and the preparation of forecasts; how to monitor and maintain quality; how to benchmark and improve; and how to stay within what can be done and yet keep your customers, who want more, on board. Learn how SAS is helping them respond to these challenges. Design of experiments (DoE) allows you to get the most from your product or service by efficiently exploring the opportunity space represented by the ranges of input settings. This knowledge allows you to produce outcomes that are more favourable to you and your customers, whatever that means in your specific context.

Using real-world situations, this talk will show how experimental designs can easily be created, analyzed and optimized. We will focus on the JMP Custom Design Platform, a powerful and flexible tool that allows building experimental designs that are specifically tailored to your needs and constraints. The Danish alcohol market has three types of alcohol: Beer, wine and spirits. The market share of wine has doubled over a 25 year period, while the market share of beer has been declining and the market share of spirits is generally low and fluctuating. In recent years the trending behavior has however changed, most likely because of changes in taxation on spirits. In the paper these market shares are analyzed by unobserved component models using Proc Ucm as models with time varying trends etc. and are well suited for this type of data. In Denmark the relative prices for the three types of alcohol have changed radically because of changes in the taxation and hence the relative prices provide good independent variables in regressions. In SAS 9.2 Proc Ucm has been extended with a Randomreg statement that allows for the varying regression coefficients. One result is that the effect of relative prices among the three types of alcohol has decreased, probably because of the increasing wealth in Denmark. The London Fire Brigade handles around a quarter of a million emergency calls a year and attends 145,000 operational incidents, of which more than 30,000 are fires. But the UK fire service is much more than a reactionary, responsive organisation. As well as a clear duty to respond quickly, effectively and professionally to all emergences, there is also a new duty to prevent those emergencies from happening in the first place.

Most of the fires that cause injury, and at times death, happen in peoples' own homes and usually as the result of a careless accident. The London Fire Brigade believes that these fires are preventable and so stopping these fires is a corporate priority.

With over 7.5 million people in London living in more than 3.2 million homes, identifying where fire risk is the greatest and who it's mostly likely to affect is an important challenge.

In his presentation Andy Mobbs, Risk Information Manager for the London Fire Brigade, will explain how its use of analytics has addressed the prevention challenge. He will show how the London Fire Brigade has used regression modelling, lifestyle profiling data and GIS tools to understand fire risks and present that information in an accessible way to the firefighters who protect the UK's capital city. The goal and focus of the Early Warning Project has been to find early indicators of potential trouble or sub¬opti¬mal performance of a complex industrial process. If such indicators are found, there may still be time to perform the appropriate actions in order to stabilize the process and avoid the problem. An alert service should pick up on a signal of upcoming trouble as early as possible, and maintain an "alert degree" which varies according to new evidence coming in. Such functionality will be more potent ? and complex ? than fixed value-based alert generation that are in abundance today.

One test case will be presented, namely the performance of the low-pressure separator on ConocoPhillips' Ekofisk 2/4 J (EKOJ) platform in the North Sea. This is a highly complex process and there are a lot of influences that are not fully understood. The EKOJ receives streams also from other platforms, meaning that performance issues on these, "up-stream" flows may influence the performance on EKOJ or, more specifically, the separator that we focus on.

A method has been designed that, given a set of sensory data (time series showing temperatures, pressures, flows etc.) around the industrial process (on EKOJ itself as well as upstream to EKOJ), extracts knowledge in the form of sets of simple rules. Such rules are intuitively understandable to the experts, but may also be used for real-time monitoring of the working system. Each rule provides diagnostic capability but may also be combined with others to yield combined evidence and an overall probability of imminent problem or failure. Rules may also be used to directly diagnose the situation at hand. Selecting customers for sales activities based on analytics is one of the most popular applications in analytical CRM. However, customers are not just waiting for their banks to contact them. First and foremost, they use ATMs, the Internet and telephones to handle their banking transactions, to gather product information or to get in touch with their advisers.

How could Erste Bank take advantage of its customers' activities and at the same time offer the right products/solutions to the right customers? Why couldn't we offer products our customers really need via those points of contact they actually use?

This presentation shows how accurate, individualized product offers for customers are determined and how the paradigm shift from a product-centered to a customer-centered view was implemented at Erste Bank. UNIQA is the largest insurance company in Austria and active in more than 16 European countries with more than 7 million customers. They have established an international, comprehensive and standardised SAS database architecture that enables an integrated data management process.

This database is not only the basis for the data mining process with SAS Enterprise MinerTM but also the central basis for the whole campaign management for the client reporting system. It is the basis for all actuarial analyses and an important source for the SAS balanced scorecard.

This presentation shows how data mining and customer data management trigger the client-oriented strategy of UNIQA. The main results (individual client ratios) of the data mining and scoring process are implemented in the UNIQA client information system (all employees have access to this system). This presentation will concentrate on the most important ratio ? customer-profit-value. This value is a prognosis of the expected client contribution margin for the next 12 months. The customer-profit-value ratio enables UNIQA to differentiate between good and bad customers and allows a very precise value prognosis not only on an individual basis but also on a segment level. This presentation will also show how UNIQA uses these customer-oriented scores to steer the sales force. In recent years an amazing growth in analytic technology can be witnessed. Corporations across several industry verticals have recognized the value analytics can provide in not only driving compelling business solutions, but also when helping differentiate themselves to customers, investors and regulators. Most importantly, these corporations have established analytics as a critical differentiator in providing both business and customer value across every customer interaction and business process. We believe that the future of business analytics will be mainly driven by the amount and type of data available: business process data, data about people (for example social network data), sensor data and unstructured data (text and video, particularly from the web). The area of business analytics will be faced with solving key challenges using these new kinds of data sources: acting on new data at the right time (real time), increased need to create and maintain predictive models, and exploiting unstructured content and semantic integration. We will discuss these challenges and we will provide an outlook on how SAS is going to address these challenges. An analysis of the full text of all occupational injury claims associated with threats or violence among Police Officers and Security Guards reported to the Swedish National workers' compensation insurance 2004-2007 was undertaken. The text-mining analysis has generated clusters of details on hazardous exposures and accident processes which point to practical modifications in training and professional operating procedures.

The analysis of free-text reporting of traumatic injuries, accidents and incidents promises to better represent the dynamics and detail of exposure and accident process and, when applied to valid and representative insurance data on occupational injury with the help of suitable text-mining software, will provide industry groups and local companies with decision support for prevention of safety management. For a long time the traditional approach to measuring forecasting performance in a business context has involved a narrow focus on analyzing percentage errors, meanwhile it has been well known that this, for a number of reasons, can be problematic.

This presentation shows how it is possible to move from the traditional approach, where there is a narrow focus on percentage errors, to a more analytical rewarding focus on scaled errors. In the case of the LEGO Group this change was motivated by the need for a performance statistic that would be useful for effective internal benchmarking and has lead to the development of the LEGO Forecasting Performance Index (LFPI), a powerful measure which is based on the latest developments in forecasting research.

Lauge Valentin will present benefits of moving beyond MAPE, establishing a forecasting improvement process for continuous improvements, and show how to conceptualize a radically different forecasting performance measure so it becomes management friendly. As a major European energy provider, Nuon focuses on customer value and sustainable energy sources. Customer insight drives their total business from strategy, objectives, brand image and marketing, and helps them gain market share continually. Nuon is currently evolving from a mass-marketing approach toward a more targeted, customer-specific approach. Through in-depth customer insight analyses they are trying to implement an efficient segmentation strategy, which is a serious challenge to many companies. Nuon is dealing with huge amounts of data and finding the valuable knowledge of what drives certain customer segments can be tricky. This presentation will focus on the development of internal credit scores in order to reduce bad debt risk within the company and how these results are implemented into daily business processes and strategy. Classical Six Sigma tools such as Quality Function Deployment, Measurement Systems Analyses, Design of Experiments and Statistical Process Control are known to be able to significantly reduce Cost of Poor Quality if C_p is lower than the Six Sigma target of 2. The pharmaceutical industry has on average a C_p level of 1 partly due to mixing up specification limits and control limits. The benefit potential is therefore huge from applying these methods within this industry. Adaptation of the Six Sigma tools is a natural first step on the pharmaceutical industry's travel from Quality by Inspection and Scrap to Quality by Design. The presentation will present the tools and show examples on how they have been used within the pharmaceutical industry. JMP has been a good software choice for this work. It has the necessary analytical methods as standard, good visualization properties and is easy to use for nonstatisticians. The usefulness of predictive models depends on several factors, but it could be argued that predictive performance and acceptability are the most important criteria. However, a trade-off seems to exist between these two concepts, as highly complex predictive models do not always offer good interpretability, and may prove hard to "sell" to marketing or general management. In this study, we offer the audience a number of clearly defined paths (driven by SAS software) toward improving the face validity of a predictive model, and we illustrate the trade-off between performance and interpretability using real-life case studies.