• HOME
• SPEAKERS
• ABSTRACTS
• AGENDA
• TRAINING
• SPONSORS
• HOTEL AND TRAVEL
• INTEREST LIST
• Register Now!

Enter a friend's e-mail address:

Download our brochure (.pdf).

Abstracts

This page is updated as we receive speaker abstracts, so visit often. This presentation touches on the theory and methodology to set statistical safety stock target based on forecast performance. Furthermore, independent of the method used to establish the safety stock levels, the presentation outlines the process used to monitor the efficiency of each of the targets as well as the overall inventory level. Finally, it outlines simple KPI's to communicate the results of inventory levels. Most mathematical forecasting techniques are based on an assumption of stationarity, the lack of which destroys the theory on which the methods are based and often makes forecasts look unreasonable. On the other hand, time series for which forecasts are desired often exhibit signs of nonstationarity. Reconciliation typically involves some sort of initial operation on the raw data. A series with a trend or intervention can sometimes be treated as that effect (trend or intervention) plus stationary errors but at other times, it seems that differencing is the only way to reduce the original data to stationarity. For example, most stock indices are expressed both in level and differences by the media. Indeed returns on investments are often calculated as a difference in log transformed prices prior to analysis. There are some disadvantages to differencing and one might want to avoid unnecessary differencing, or "overdifferencing". This talk will explain what is meant by stationarity, and hence nonstationarity, showing some illustrative examples. It will proceed to look at intuitive graphical approaches to detecting nonstationarity as well as more formal statistical model based tests for stationarity. The focus will be on a popular test available in several SAS products such as PROC ARIMA. The excitement over reporting tools is cooling. Executives can now answer "what happened yesterday?" with ease. But these tools have blurred the distinction between "Reporting" and "Analytics", with many people perceiving no difference what-so-ever. With data volume once again becoming our enemy, we need to find methods quickly for gleaning information from large mounds of data. In this talk, I will discuss statistical and technological methods that simplify your analytic workday and allow enable true business intelligence. From Maximum-Likelihood methods to novel uses of your database's OLAP functionality, the proper mathematical toolbox takes you from efficiency-obsessed to intelligence-minded. Firms in capital intensive industries such as high tech, biotechnology, and pharmaceuticals must make planning decisions in an environment fraught with uncertainty, increased product complexity, increased customization pressure and ever shortening product life cycles. One of the latest methods being introduced in the planner's toolbox is Prediction Markets. Predictions Markets is a cost effective way to consolidate expert opinion in a flexible format. While prediction markets have been used for many years in forecasting such things as presidential elections, only more recently have these methods been evaluated in their application to industry and demand forecasting. Participants will learn how to incorporate Prediction Markets into existing forecasting processes, how to secure participant and stakeholder buy-in, review associated forecast metrics (FVA based), and review three high tech cases. Accurate forecasts are crucial to successful planning in most organizations. However, a survey of 120 US forecasters and study of forecasting in four large UK companies suggests that many companies are not making the best use of their forecasting software. Managers in some companies used data histories that were too short to give the software a good chance of producing reliable forecasts. They also intervened too frequently in the model fitting process and were over zealous in their application of judgmental overrides of the software's forecasts. Many organizations did not monitor the accuracy of these overrides or record reasons why they were made. This talk will indicate the situations when judgmental changes to forecasts are likely to be appropriate and when they are likely to waste management time or reduce forecast accuracy. It will also demonstrate a number or methods that are likely to make judgmental adjustments more accurate on those occasions when an intervention is desirable. Like S&OP, CPFR is not a new concept, although it has been somewhat slower to take root in many companies. As S&OP matures, however, it necessitates that CPFR becomes an integral part of the planning process. Collaboration being the first "C" in CPFR underscores the importance of collaboration in that process, however, S&OP which in many companies has become consensus driven tends to be much less collaborative in the truest sense of the word. Companies that embrace CPFR fully as the next evolutionary step of S&OP will find that the old rules of the consensus process don't apply here. The LEGO Group abandoned percentage errors when evaluating forecasts and replaced them with scaled errors. The shift from percentage errors to scaled errors was motivated by the need for an accuracy statistic that would lend itself to benchmarking across product groups and be useful when evaluating forecasting performance. This led to the LEGO Forecasting Performance Index (LFPI), a powerful measure that adds value to the forecasting organization and enables effective improvement processes. Demand forecasts are a key enabler for supply chain processes. Traditionally, forecasting processes and applications are applied in CPG, Retail or other similar industries. This presentation will demonstrate how they may be utilized in non-traditional environments. We will discuss how Starbucks applied knowledge from the forecasting discipline to support both new and existing store development. Emphasis will be placed on the need for discipline, forecast collaboration, measurement and the ability to leverage data from multiple sources to ensure actionable forecasts are developed. How to best measure forecast accuracy and other performance factors remain key issues on managers' minds. Many companies often pursue forecast accuracy levels deemed by senior management without consideration of realistic benchmarks. Published and unpublished sources of forecast accuracy are questionable because they may apply to very divergent market segments, company structures and industry types. This presentation will discuss the ways in which forecasting performance can be evaluated, with implications posed for each. In addition, the topic of "forecastability" will be discussed – a topic which lately has been adopted by companies in the course of establishing internal forecast benchmarks. Company case studies will be presented for the purposes of generally illustrating considerations and issues surrounding forecasting performance measurement. An enormous amount of recent attention has been directed at forecasts leveraging Unobserved Components Models (UCM). This is particularly exciting for analysts working with retail scanner data. Forecasting practice with scanner data has been dominated by regression-based models for decades since users need to assess causal relationships with marketing efforts like promotions and pricing. UCM take this to the next level by decomposing market response into both time series components and the regression effects of a predictor series like promotional activity. Analysts can build their UCM with maximum-likelihood estimated parameters and then optimize fit by selecting various interventions, level shifts, autoregressive components, or exogenous variables to explain the multivariate processes inherent in the data and the marketplace. For Consumer Packaged Goods (CPG) manufacturers and retailers this provides immediate insights that both explain and predict consumer behavior. The outcome is not only strong forecasts, but results that ensure buy-in and subsequent action by users. This discussion will provide an overview of the methodology, a list of software options, and a case study with validations demonstrating the power of UCM to "raise the bar" for forecasters leveraging scanner data to predict and obtain future demand.

Firms in capital intensive industries such as high tech, biotechnology, and pharmaceuticals must make planning decisions in an environment fraught with uncertainty, increased product complexity, increased customization pressure and ever shortening product life cycles. One of the latest methods being introduced in the planner's toolbox is Prediction Markets. Predictions Markets is a cost effective way to consolidate expert opinion in a flexible format. While prediction markets have been used for many years in forecasting such things as presidential elections, only more recently have these methods been evaluated in their application to industry and demand forecasting. You will learn how to incorporate Prediction Markets into existing forecasting processes, how to secure participant and stakeholder buy-in, review associated forecast metrics (FVA based), and review three high tech cases. Sales and Operations Planning (S&OP) processes have been around for twenty years or more, with most companies having implemented them a long time ago. However since then, supply chains have become more complex involving global sources & markets, longer supply lines, and more short product-life-cycle products. This has led to a resurgence of interest in S&OP with many companies revising their processes or just starting them. This talk will discuss the basics of the process and the improvements that are needed to manage today's complex, global, and dynamic supply chains. Companies spend a great amount of resources in developing new products. Predicting their demand is a critical part of the planning process. It is also one of the most challenging tasks for forecasters. By definition, sales data history does not exist for new products. Traditional time series methods of forecasting do not apply, and the usefulness of new product diffusion models is limited when little or no data are available. In such a case, the forecasting task is often delegated to one or more experts who base their judgment on the performance of past similar products.

We propose a flexible structured judgmental procedure composed of four main steps, whereby the expert's judgment is aided by statistical tools made available by a computer system. At each step, the expert can interact with the system and modify the suggestions provided by the statistical analysis. In the query step, candidate analogous products are identified among existing products by selecting attributes that match the new product. In the filtering step, the candidate products are further refined into a set of surrogate products by clustering similar series according to a similarity measure. In the modeling step, several statistical models are fitted to the surrogate products data. Finally, in the forecasting step, forecasts are generated from the fitted model, and adjusted by the expert to account for factors not considered by the models. As sales data for the new product become available, they can be incorporated into the process to monitor the performance and further refine the forecasts. Business organizations collect large amounts of time-stamped data related to their suppliers and/or customers through web sites or other transactional databases. A business can have many suppliers and/or customers and can have a set of transactions associated with each one. Mining these time-stamped data can help business leaders make better decisions by enabling them to listen to their suppliers or customers via their transactions. Each set of transactions can be large, making it difficult to perform many traditional data mining tasks. This paper proposes techniques for large-scale analysis using similarity measures combined with automatic time series analysis and decomposition. After similarity analysis, traditional data mining techniques can be applied to the results along with other profile data. This paper demonstrates these techniques using SAS/ETS, SAS/STAT, SAS Enterprise Miner, and SAS High-Performance Forecasting software. Most companies' supply networks are their next competitive advantage opportunity. Managing thousands of suppliers with thousands of SKUs is a major coordination and planning challenge. Some estimates of the coordination costs (and impacts due to poor coordination) are approaching 30% of the purchase costs of supplies. Developing integrated demand and operations planning with a company's supply base is a way to attack this opportunity. This presentation discusses recent research and case studies on attempts to address this opportunity. The problem of forecasting intermittent time series (time series that contain a lot of zero values intermixed with non-zero values) is very challenging, and the importance of finding appropriate error measurements for this class of problems should not be overlooked. In Revenue Management applications, intermittent time series often occur when estimating the potential value of inventory. In this case it is important for forecasts to accurately predict zero values versus non-zero values. For this type of forecasting problem, forecasts performing well on some common metrics such as MAPE and RMSE are often less useful than forecasts that have poor MAPE characteristics, but have good Receiver Operating Characteristic (ROC) curves. In this presentation, we discuss different methodologies for forecasting intermittent time series in addition to ways to evaluate their forecast applicability. The product life cycle and forecasting approaches from pre-clinical compounds to post patent strategies will be discussed. This talk will explore epidemiological models (when to use and when not), dynamic patient flow models, head to head models, composite share stealing models and statistical models using time series data. We present a technique for improving a time series forecast based upon a partial realization of the variable that we are forecasting. In industries with perishable assets, there is typically a horizon during which demand for a product is being accumulated. For example, in the lodging industry the products under consideration are room nights at a lodging establishment. Demand for these room nights accumulates over a booking horizon. As the consumption date approaches, we accumulate more information about the final demand. Traditional time series methods do not use this information in their forecast. In this paper, we describe a method which updates a time series forecast by taking realized bookings into account. This technique is based upon Bayesian principles. Specifically, Monte-Carol simulation is performed to estimate the conditional expectation of the quantity being forecasted. The variables that we condition on are the realized bookings to date. Empirical studies show that incorporating this information improves the accuracy of the forecast. The presentation explores the challenges presented by very seasonal and cyclical sales patterns. Keys include examining the seasonal characteristics of each product or groups of products individually, utilizing the knowledge of the marketing and sales organization and matching the appropriate statistical tools to the sales data. This presentation will review the importance of building alliances to improve the understanding and acceptance of forecasting. Accuracy is often the focus of an organization, but other factors may be equally important to a successful forecast organization. The development of a good statistical demand forecast is an interesting technical exercise that often will give an organization insight into customer demand, a fundamental driver of their business. Beyond this, why do we need a good demand forecast? What do we do with it once we have expended the often significant effort to create one?

The key to realizing maximal benefit from our demand forecast is to tightly integrate the forecasting process within the organization's planning processes. These planning processes are the mechanisms that an organization uses to marshal its resources to achieve the desired goals and objectives. A demand forecast at the required level of detail is a core input to these planning processes.

In this talk we will explore where and how we use a demand forecast in various planning processes. In their 2007 report, AMR suggested that one of the best ways to improve forecast accuracy is by using downstream data from distributors and channel partners more effectively. Distributors often have first hand information about the market dynamics for the products they sell. This point-of-sale (POS) or sell-through data can reveal patterns and insight into customer behavior as well as future product demand. However, this critical information is often disregarded or unavailable to upstream functions in the supply chain, resulting in a distorted view of true customer demand.

In this session, listeners will learn

• methods for leveraging sell-through data in forecast models to improve overall sell-in forecast accuracy
• applications of these methods in practice
• the value of information sharing in the supply chain and how it can lead to more effective forecasting and demand planning

Managerial and quantitative forecasts each have their own strengths and weaknesses and can bring different information to the forecasting process. This session will address different ways of combining managerial and quantitative forecasts and the advantage and disadvantages of each approach, factoring in the inherent biases of each method. Dr. Sanders will also discuss principles that have been developed for deciding when and how to use managerial judgment in adjusting quantitative forecasts, based on real-world experience and extensive research of prominent business researchers. When an organization is planning to incorporate forecasting within their business, they will often look at pre-production evaluations including proof-of-concept (POC) and pilot runs. The POC enables the organization to determine whether their existing data can support forecasting. However, the pilot provides the information necessary to determine if the POC would extend to a real operational environment by applying forecasting to a subset of operations. The challenge with a pilot is that it has to be close enough to real operational conditions to provide an effective demonstration of the forecasting impact on operations. To be successful with a pilot, there are several considerations that must be addressed within the project, including: pre-pilot planning and code development, pilot live operation, and offline analysis. The outcome of a pilot needs to be sufficient information to determine the value and cost of a full forecasting deployment. This presentation will describe the approach taken on a successful forecasting pilot project for a client.

---