Abstracts

This page will be updated as we receive speaker abstracts, so visit often. The following abstracts have been provided to-date:

Keynote Addresses

One way to forecast is with quantitative methods. These involve models fit to data. The models forecast future values of a series in terms of correlation patterns picked up from the historic data, and can also use information from predictor variables. I will review basic ARIMA models then look at examples of transfer functions and intervention analysis as well as ARCH models in which local volatility is mathematically modeled. These methods go from fairly simple to quite sophisticated and are all incorporated in SAS software. Statistical methods and expert management judgment can bring complementary benefits to the forecasting process. For example, statistical methods are superior at finding patterns in large volumes of data, while management judgment can play a useful role in taking into account one-off special events. Despite its potential value, judgment is also subject to both inconsistency and a number of widely documented biases.

Drawing on a major study of forecasting in supply-chain companies, this talk will first look at the biases that are associated with the use of management judgment in forecasting. It will then demonstrate a range of methods that are designed to allow judgment to play an effective role when it used in combination with statistical forecasts. A practical perspective on extending Revenue Optimization to supply chain management. The presentation will focus on synergy's within the revenue optimization field and implications to supply chain management. SAS is the most widely used statistical/data mining package in the world. Teradata is the world's fastest and most scalable data warehouse. Each, individually, can handle any data-driven business question; working together all is possible at incredible speeds. The partnership between Teradata and SAS is strong and thriving. Warner Home Video is one company leveraging that relationship by making effective use of the power of both by exploiting the strengths of both. This presentation will review the approach being used by business analysts at Warner Home Video in implementing statistical/predictive models while utilizing Teradata to do the heavy lifting and using SAS's rich statistical features gaining order of magnitude in performance as compared to running the total statistical modeling activity purely on SAS servers.

Track One: Forecast Data Considerations

The healthcare industry is just now beginning a transition from paper to electronic record-keeping. Integration of different systems remains a problem. In addition, the definition of customer and supplier are fuzzy and can change based upon decision point. With time stamps provided by electronic medical records, it is possible to use transactional time series methods (PROC HPF) to examine efficiency in providing timely patient care and reducing waiting for patient beds. The structured use of hospital facilities such as clinics and operating rooms can also be examined with optimization and time series. However, because of the lack of integration between hospital units, considerable pre-processing of data is required to merge time data. Preliminary results demonstrate that bottleneck times in the hospital emergency department can be relieved by staggered scheduling of staff. Waiting time for a patient bed is dependent upon discharge times, and optimization techniques should be used for OR scheduling. Many companies operate in an increasingly competitive environment and accurate demand forecasting is crucial to their success. For these companies, marketing actions, such as promotions, and external events, such as the weather and holidays, have a major influence on demand. This presentation, which is based on extensive research in supply-chain companies, will examine a variety of approaches to the problem of incorporating these effects into the forecasts.

The key influence on the approach to adopt is the availability of data. If data have been kept on the major events that effect demand, statistical model-based approaches will provide the best approach. Examples of how to incorporate complex weather effects in both retail and beverage forecasts will be discussed. More commonly, either the event history has not been properly recorded or the expected future event is not thought to be similar to those seen in the past. Here management judgment must play a major role. But such judgments do not always lead to improved accuracy beyond that delivered by the statistical forecasting software, as the case evidence we have collected shows. Often adjustments are even made in the wrong direction! It's going to be 'good'; but it turns out worse. The presentation concludes with a discussion of software enhancements that might help company forecasters improve their estimates of the impact of events. Dynamic changes, both inside and outside your organization, require accurate, flexible, up-to-date forecasting to optimize your supply chain.

Finding the best analytical models for your numerous business lines and product families can be tedious. Nowadays, optimizing your inventory means getting an accurate picture of every item as well as reconciling forecasts performed at various levels of your product hierarchy. For large scale environments, this could yield millions of daily forecasts. A robust and scalable, yet nimble forecasting solution is needed to address these needs.

In turn, getting the most from forecasting technology demands an equally robust, scalable and nimble IT infrastructure. This session will help you get prepared for the most complex forecasts and modeling your Financial Planners and Statisticians can dream up. We will present a sizing model, based on empirical data that will help you identify the system elements you will need to handle a range of forecasting requirements. Typical configuration examples will be provided to help you plan for implementing a forecasting solution. We will offer best practices suggestions for setting up an environment to help you get the most performance out of your system.

And as your business requirements scale, we will offer a set of guidelines on how to scale your infrastructure to meet those demands.

Organizations are continually trying to extract more information from the large volumes of time-stamped data they collect in order to become more agile and operate more efficiently. The most obvious use for time-stamped data is forecasting. Over the past few years tools have been introduced which vastly improve many large scale forecasting problems via sophisticated automation. Some forecasting and planning problems still remain inadequately addressed. In these areas, time series data mining and visualization can be used to augment traditional time series techniques.

Track Two & Four: Forecasting Approaches

A technically sound forecast goes nowhere if it is not accepted by those in power, whether they are politicians or upper management. Often the forecast gets manipulated to satisfy political ends or meet targets and plans. This presentation tackles some tough issues that organizations face, such as:

• How can we minimize organizational biases that impact forecasting?
• How can we improve collaboration among players with conflicting interests?
• How can we achieve buy-in or acceptance of the forecast by those in power?
• What's the best organizational design for a neutral forecast process?

We will discuss six simple strategies to overcome these problems and allow the technically sound forecast to prevail over politics. Essential steps in making business forecasts and the evaluation of business forecasts are addressed. A detailed look at U.S. Coast Guard Aviation's logistics supply chain and their approach to forecasting intermittent demand. The Coast Guard's aviation parts inventory is comprised of numerous parts that experience intermittent demand and long lead-times (both of which are plagued by large variability). The ability to accurately forecast this intermittent demand over the entire period is paramount to building more accurate budget models and developing procedures for the timing and size of replenishment orders. Utilizing historical demand data and a pre-determined service level, the Coast Guard embraces the tails of the demand distribution to better capture the true demand behavior of each part in their inventory. The ultimate goal is to reduce costs while maintaining a high service level. There are four steps to modeling: using the right data; selecting the right model; overlaying the right judgment over statistical forecasts, where necessary; and monitoring and updating forecasts. In this session, we will discuss all four issues.

We'll also learn what to look for in your data, how to fix potential problems, and how to use data patterns to select the right model. In addition we'll discuss some non-traditional forecasting models that are very simple, yet work very effectively, as well as learn some good strategies for monitoring forecasts and how to deal with products that are difficult to forecast. We'll conclude with a survey of the various forecasting models that are used in business, based on the survey conducted by the Institute of Business Forecasting. Accurate long range (1-2 years) customer churn forecasts are essential for strategic planning and optimizing capital for retaining and growing the customer base. While adapting survival analysis for business forecasting is a promising approach, applying it to large scale real world data presents many challenges, especially in representing business policies in modeling assumptions. In this talk, we will look at case studies of such challenges and their resolutions, primarily from the Telecom area. Integrating consumer demand to improve shipment forecasts has become a high priority in the Consumer Packaged Goods (CPG) industry over the past several years. Until recently many factors, such as data collection and storage constraints, poor data synchronization capabilities, technology limitations, and limited internal analytical expertise have made it impossible to either "integrate with consumer demand," or "link to shipment forecasts." With improvements in technology, data collection, and storage, along with improved analytical knowledge, CPG companies are now looking to integrate consumer demand with their shipment forecasts to capture the impact of marketing activities on shipments. As a result, a technique called "Multi-tiered Causal Analysis (MTCA)" is receiving renewed interest. MTCA is not a technique but rather a procedure or process that considers marketing and replenishment strategies jointly rather than creating two separate forecasts (i.e., one for consumer demand and another for factory shipments). Businesses spend lots of time and money on forecasting, yet are frequently dissatisfied with the results. While no method can guarantee the level of forecast accuracy achieved, it is possible to control the process used and the resources invested. The Lean Approach improves forecasting process performance by identifying and eliminating process waste. This presentation offers methods for tracking the "value added" by each process step and participant. Those steps and participants that are failing to make the forecast better -- or may even be making it worse! -- can be identified and safely eliminated. The result is a streamlined forecasting process that delivers quality forecasts as efficiently as possible.

Track Three: Forecasting Applications

A retail sales data set is recorded hierarchically for corporate-distribution center-SKU. Accurate forecasts are desired at each level of the hierarchy. Top-down, middle-out, and bottom-up reconciliation methods are applied to the data to investigate the merits of each approach. Issues related to large-scale forecasting of retail data are discussed. Energy utilities rely on sales forecasts to set rates and plan for adequate supplies. These business functions often take place in an environment of regulatory oversight that demands transparency and professional rigor. This presentation explores the approach used to forecast natural gas demand for a northwest utility. Econometric models that combine weather and economic terms were developed to forecast residential demand for natural gas. The specification of the weather term in econometric models has important implications for forecast accuracy. The use of SAS to optimize the specification of the weather term as well as estimate model coefficients will be discussed in this presentation. Electric and gas utilities have traditionally conducted load forecasting to plan capital expenditures to meet long-term load requirements and provide weather-normalized energy sales and revenue forecasts for near-term operational planning and budgeting. In an industry where being within +/- 5% was at one time acceptable, today's utility executives desire that "backcast" errors are no larger than 1%. Moreover, the traditional "top down" aggregation into residential, commercial, industrial, and governmental customer sectors no longer meets utility business requirements.

Today's utility executives want to know considerably more. Where is load growth expected to occur? How much do growth rates vary across our territory? What kinds of customers are contributing to growth? What about in areas with declining energy sales? What can we do to meet customer needs and better manage growth? These questions are akin to questions posed by manufacturers and retailers, and demonstrate that many issues related to forecasting in a "just-in-time inventory management" world have now reached the utility industry. However, the ability to provide this type of detailed, disaggregated forecast information required an unacceptably large staff of forecast analysts.

Until now. We combined traditional top down econometric load forecasting techniques with the state-of-the-art ability of SAS Forecast Server to quickly address thousands of related, disaggregated series in what we call Multi-Dimensional Forecasting (MDF). Our presentation will describe the MDF process and show how this new approach to sales forecasting – where the successes of the past are combined with the technology of the future – can answer today's business forecasting questions.

In Deregulated Electric Markets, controlling cost is critical to assuring company profitability because there is no guaranteed rate of return as is the case for a regulated utility. Accurate load forecasts are key drivers for controlling cost since fuel costs are the single largest cost component tied to running an electric utility. Bilateral contract positions for fuel for self owed power generation or for purchased power from other utilities are constantly under evaluation. Knowing the open position in relation to the forecasted load determines long term or short term buy or sell conditions in an ever changing market improves the likelihood of supplying load at lowest cost. Fast moving weather fronts, extreme weather, new or lost contracts, mass customer switching, market price fluctuations and the regulated rules for financial settlement make the forecasting process very complex, and present a dynamic set of challenges for the forecaster. The presentation shall provide some insight into the complexity of this forecasting activity. The eBay community of over 192 million users makes for IT capacity management challenges unlike any other. How do we make the right decisions for scaling, performance, and availability while managing an aggressive IT budget? Through the development of eBay's performance data warehouse, accurate forecasting models, and timely reports, eBay's Capacity Management teams have developed a capacity management practice that brings together various business, system, and infrastructure metrics. Through these practices, by delivering timely and accurate capacity forecasts, and by working with technology organizations throughout the product development lifecycle, eBay's Capacity Management team is able to accurately predict capacity needs for the "the World's On-Line Marketplace."