Abstracts

We are continuing to received abstracts from our speakers for F2009. Please check back often for updated information. This is a case study of a mid-tiered, regional CPG company who decided to reposition their company in the marketplace, roll-out with national distribution for their products, and realized that they needed a more robust demand forecasting and planning solution to help manage this growth. In this discussion, we will talk about how Marcal approached this process readiness, and plans to implement a new demand planning solution. It will also discuss the challenges of change management. Composite forecasting, introduced into economics and business literatures by Bates and Granger (1969), recognizes that policy planning requires a forecast of policy relevant variables and that no one forecast (model or person) contains all available useful information. Early work on composites applied this idea to formal model-based forecasts (e.g., two alternative ARIMA representations of the same data set). In agriculture, as well as business in general, much potentially valuable forecast information exists as informal models. These are based on subjective expert opinion (human judgment). Bessler's work in the early 1980s looked for methods to combine formal model-based forecasts with subjective expert opinion. The evidence presented in Bessler and Brandt (Applied Economics 1981 and AJAE 1981) shows clearly that such combinations work. Bessler has recently focused his attention on the dependence among forecasts which is of a causal nature. That is to say, he investigates whether one can use the correlation structure among the individual forecasts and infer the causal structure that lies behind the data.

In the June SAS address, Bessler will discuss algorithms of inductive causation by Judea Pearl (Causality, Cambridge 2000) and their use in exploring the correlation structure on alternative forecasts. Bessler's example considers forecasts from an ARIMA model of historical time series and from two subjective experts. These three alternative forecasts have been discussed in earlier studies by Bessler in the Journal of Forecasting 1983 and in the Journal of Economic Behavior and Organization 1992. Results of Bessler's current work will be presented and will show that forecasts from the ARIMA model serve as an information source (root cause). One subjective expert forecast series is found to be a mediating cause and the other expert forecast is an information sink -- the latter providing no new information on the eventual outcome of the random variable of interest. Extensions to composites of probability forecasts are considered. Have you ever been caught in a situation where most of the really good data for a forecast came from sources that were more qualitative than quantitative? At Novozymes one of the fastest growing sectors of the business was the Fuel Ethanol industry. While the existing infrastructure was well known and the demands of these sites were also general knowledge, the changes that were on the horizon were completely unknown. From changing government mandates, to the size of the US corn crop, to the biggest portion capacity under construction or in talks to be built. Understanding how to use these unorthodox data streams, in conjunction with hard data is the crux of this discussion.

In my presentation, I will show:

• A real world example of how an ever changing environment in a complex biotech industry was dealt with, and some practical modeling examples.
• Tips on how to deal with the qualitative data in as practical a method as possible.

What an attendee will learn from my session:

1. How to rationalize qualitative data, in order to use it consistently in a data model.
2. How to move an organization away from point forecasting to range forecasting, and begin to have them work with this new dynamic.

There is considerable evidence that states the macro-economy affect the chances that a consumer will default on one or more of his/her consumer loans. This has considerable implications for lenders who wish to predict the probability of a borrower defaulting if they are given a loan, and for a lender who wishes to manage a portfolio of loans and is interested in the losses it may incur in an economic down-turn and the value at risk of such a portfolio. In this presentation we explain how macro-economic variables can be incorporated into credit risk models. We will first explain how continuous time survival analysis can be used to explain and predict when a borrower will default in the next time period, conditional on not having defaulted before, and we will give the results of a parameterization of such a model. Secondly we will explain how one can incorporate behavioral (time varying-individual specific) variables into a discrete time survival model and show the implications of stressing variables in the model for the distribution of losses in a portfolio. We construct a framework for high-frequency nowcasting of business conditions, potentially in real time. We use a variety of stock and flow data observed at mixed frequencies, and we use a dynamic factor model that permits exact filtering. We discuss a number of issues and open questions that have arisen in the real-time use of our methods at the Federal Reserve Bank of Philadelphia. Many business functions can benefit from appropriate forecasting techniques. However, most companies cannot afford to staff each department with statisticians and actuaries to create sound forecasts. Our presentation will look at how a collaborative, enterprise approach can bring value to technical and non-technical areas alike. We will provide an example that focuses on a recent effort between several internal State Farm departments, SAS consulting, and the University of Illinois.

Initial forecast models are proposed by key analytic areas within the company. Model development is a collaborative effort relying on expertise inside and outside the company. Once models are tested for appropriateness, business context is provided around the use of the models, and the models are then made available to other areas.

Models are stored in templates that can be updated and distributed in Excel using SAS AMO. Models used in non-technical areas have been pre-selected by subject experts as appropriate for particular types of data. The template approach allows the subject expert to control the models being used, limiting the models to those producing the best results. Survey redesign and other changes in the methods of repeated surveys can introduce breaks in the related time series. In an ideal situation, a parallel run is performed to estimate said break and to decide if a linking or bridging exercise is required. When a parallel run is not feasible and segments of time series which share no explicit overlapping period must be linked, a common reference period is established with forecasts. In most cases, aggregate consistent forecasts and linked series are expected. Hierarchical forecasting can be used to restore coherence in set of time series subjected to simple tree-like aggregate structure. Generalized methods such as multi-dimension reconciliation are used for more complex aggregate rules.

Presented here are two detailed examples of linking non-overlapping times series from Statistics Canada: linking evaluation for the Industrial Consumption Energy Survey (ICES) and bridging the Canadian Travel Survey (CTS) to the new Travel Survey of Residents of Canada (TSRC). Automatic forecasting tools such as SAS Forecast Studio® and X-12-ARIMA are used, compared and complemented with a reconciliation strategy when necessary. In today's times of frequent global economy fluctuations and highly volatile weather conditions, managing risk within an energy company is an ongoing challenge. This presentation will show how Old Dominion Electric Cooperative (ODEC) uses SAS solutions to arm its management and executive team with the needed intelligence to effectively manage its earnings - as well as its risk. An electric generation and transmission cooperative, ODEC is a power provider to its 11 member systems. The presentation addresses the key challenges faced by ODEC and its members, and how ODEC has leveraged the SAS Enterprise Intelligence Platform to meet these challenges:

• Forecast both demand and energy in a granular model that reflects its members' characteristics
• Utilize this model for short-term, mid-term, and long-term forecasts
• Achieve results within an acceptable confidence interval for managing ODEC's energy purchases
• Employ "what-if" scenarios to determine how economic and weather-related factors play into the forecasted results

After a brief review of the various forces, domestic and global, that set the stage for the current economic crisis, we will discuss the timing of a turnaround in light of the remaining headwinds. What, if anything, can we learn from previous experience and what is inherently different? Particular attention is given to the various policy efforts pursued and those options remaining. Perhaps most importantly, forecasts, statistical or otherwise, also need to confront the long-run macro-economic imbalances that represent pronounced challenges for the U.S. and global macro-economy. Few macroeconomic data series escape revision, and some are subject to large revisions over an extended period as more complete source data become available. Nevertheless, conventional estimation and forecasting uses latest-vintage data throughout. As the sample period is extended, last period's data set is thrown out and replaced by a new one. This practice is often sub-optimal, because the first-release and lightly revised data that are substituted into the estimated forecasting equation are likely to have statistical properties very different from those of the heavily revised data that dominate each sample and, hence, each estimation. Conventional practice, thus, mixes apples with oranges.

To improve on conventional estimation and forecasting one must make some assumption about the nature of data revisions. Assumptions in the literature range from the textbook measurement-error model?according to which government statistical releases equal the truth plus white-noise error?to the pure "news" model, in which government estimates make full use of available information, so that subsequent revisions are completely unpredictable.

In his presentation, Evan Koenig illustrates the pitfalls of the conventional approach to estimation and forecasting. He discusses how these pitfalls can be quite easily avoided without imposing strong a priori assumptions on the nature of data revisions, and demonstrates the impact that correct estimation can have on real-world forecasting performance. Many forecasting support systems do not enable the practice of sound forecasting principles. As might be expected, the lack of adherence to forecasting principles results in large out-of-sample prediction errors for the organization that uses these systems. Despite this obvious flaw, the software vendors that make such bad software are often successful in selling their software to willing buyers for large profits. Many members of the professional forecasting community place too much reliance on software vendors to diffuse forecasting principles. In the current marketplace, creating software that enables the practice of sound forecasting principles does not lead to software sales. The professional forecasting community needs to help software vendors who try to diffuse sound forecasting principles. This paper describes some of the software marketplace dynamics involved with the buying and selling of forecasting support systems. During this presentation we will review all key components for the evaluation process for new products. We will highlight the most commonly used forecasting approaches for new products and their practical application.

Key elements:

• Why are new product evaluations performed?
• NPV the most frequently used metric to assess new product value
• Elements of the NPV
• Key value drivers
• Why is the NPV used almost universally
• How to analyze NPV's
• How is the NPV used to drive key decisions

The SAS Telecommunications Intelligence Solution includes nearly 10 modules. In this presentation we will discuss three of them: Retention, CrossSell-UpSell, and Payment Risk as they were implemented in Orange Dominicana. We will provide an overview of the TIS product and a series of enhancements that contributed to significantly expand the scope of its modules and improve their effectiveness for forecasting. We will follow the data from its raw form to its modeling-ready transformations, logistic models development and implementation, as well as performance tracking. Let's face it: preparing data is not something a forecaster is keen on. Still we have to accept the fact that data is hardly available in the right format for forecasting. Data preparation is as important as the modeling algorithms themselves because ultimately the quality of results depends on the quality of the data. The adequate preparation of the data is an important success factor for analysis. In this presentation common challenges of data preparation for forecasting will be highlighted and some guidelines of how to address them will be shared. Today, more than ever, supply chain executives are faced with increased pressure for new value add and cost reduction vehicles. Economic conditions are reinforcing the mandate for tighter, more efficient supply chains. The next generation of business process outsourcing is focused on 'core' supply chain processes as a strategic approach to creating significant economic value. Demand Management is now center-stage as an outsourcing candidate. The risks associated with outsourcing these capabilities are manageable and the gains realized in performing these functions better than your competitor outweigh these risks. For many companies where it is not your core competency, Demand Management is too strategic not to outsource. Existing underwriting approaches have put too much of an emphasis on past loan performance and historical market conditions and not enough on the borrower's current and future capacity, and mortgage product risk characteristics. This paper introduces an effective comprehensive credit assessment framework (CCAF) that can provide early warning of risk and forward-looking analyses that do not rely on the premise that the past determines the future. CCAF is a unique rating system that extends existing credit scoring to embrace all relevant factors and business context so that lenders can classify credit risk and decision all transactions in a more, effective and transparent manner. CCAF addresses the limitations of existing lending systems imposed by insufficient credit and market information and sampling timeframes, and static data-driven factor selection and weightings. This robust and flexible approach introduces human judgmental factors into forecasting process by combining Bayesian and scenario-based methods. The end result is more accurate risk measurement for lenders and greater loan affordability for borrowers.