forecasting models.
The 2008–09 economic recession was evidence of a situation where the use of proper Xs in a multivariate in X “leading indicator” framework would have given some companies more warning of the dilemma ahead. Services like ECRI (Economic Cycle Research Institute) provided reasonable warning of the downturn some three to nine months ahead of time. Univariate forecasts were not able to capture these phenomena as well as multivariate in X forecasts.
The external databases introduced above not only offer the Ys that businesses are trying to model (like that in NAICS or ISIC databases), but also provide potential Xs (hypothesized drivers) for the multivariate in X forecasting problem. Ellis (2005) in “Ahead of the Curve” does a nice job of laying out the structure to use for determining what X variables to consider in a multivariate in X forecasting problem. Ellis provides a thought process that, when complemented with the data mining for forecasting process proposed herein, will help the business forecaster do a better job of both identifying key drivers and building useful forecasting models.
Forecasting is needed not only to predict accurate values for price, demand, costs, and so on, but it is also needed to predict when changes in economic activity will occur. Achuthan and Banerji—in their Beating the Business Cycle (2004) and Banerji in his complementary paper in 1999—present a compelling approach for determining which potential Xs to consider as leading indicators in forecasting models. Evans et al. (2002), as well as www.nber.org and www.conference-board.org, have developed frameworks for indicating large turns in economic activity for large regional economies as well as for specific industries. In doing so, they have identified key drivers as well. In the end, much of this work shows that, if we study them over a long enough time frame, we can see that many of the structural relations between Ys and Xs do not actually change. This fact offers solace to the business decision maker and forecaster willing to learn how to use data mining techniques for forecasting in order to mine the time series relationships in the data.
1.6 What is a Time Series Database?
Many large companies have decided to include external data, such as that found in Global Insights, as part of their overall data architecture. Small internal computer systems are built to automatically move data from the external source to an internal database. This practice, accompanied with tools like the SAS® Data Surveyor for SAP (which is used to extract internal transaction data from SAP), enables both the external Y and X data to be brought alongside the internal Y and X data. Often the internal Y data is still in transactional form that, once properly processed, can be converted to time series type data. With the proper time stamps in the data sets, technology such as Oracle, Sequel, Microsoft Access or SAS itself can be used to build a time series database from this internal transactional data and the external time series data. This database would now have the proper time stamp and Y and X data all in one place. This time series database is now the starting point for the data mining for forecasting multivariate in X modeling process.
1.7 What is Data Mining for Forecasting?
Various authors have defined the difference between “data mining” and classical statistical inference (Hand 1998, Glymour et al. 1997, and Kantardzic 2011, among others). In a classical statistical framework, the scientific method (Cohen 1934) drives the approach. First, there is a particular research objective sought after. These objectives are often driven by first principles or the physics of the problem. This objective is then specified in the form of a hypothesis; from there a particular statistical “model” is proposed, which then is reflected in a particular experimental design. These experimental designs make the ensuing analysis much easier in that the Xs are orthogonal to one another, which leads to a perfect separation of the effects therein. So the data is then collected, the model is fit and all previously specified hypotheses are tested using specific statistical approaches. In this way, very clean and specific cause-and-effect models can be built.
In contrast, in many business settings a set of “data” often contains many Ys and Xs, but there was no particular modeling objective or hypothesis in mind when the data was being collected in the first place. This lack of an original objective often leads to the data having multi-collinearity—that is, the Xs are actually related to one another. This makes building cause-and-effect models much more difficult. Data mining practitioners will mine this type of data in the sense that various statistical and machine learning methods are applied to the data looking for specific Xs that might predict the Y with a certain level of accuracy. Data mining on transactional data is then the process of determining what set of Xs best predicts the Ys. This is quite different than classical statistical inference using the scientific method. Building adequate prediction models does not necessarily mean that an adequate cause-and-effect model was built, again, due to the multi-collinearity problem.
When considering time series data, a similar framework can be understood. The scientific method in time series problems is driven by the economics or physics of the problem. Various structural forms can be hypothesized. Often there is a small and limited set of Xs that are then used to build multivariate in X times series forecasting models or small sets of linear models that are solved as a set of simultaneous equations. Data mining for forecasting is a similar process to the transaction data mining process. That is, given a set of Ys and Xs in a time series database, the goal is to find out what Xs do the best job of forecasting the Ys. In an industrial setting, unlike traditional data mining, a data set is not normally available for doing this data mining for forecasting exercise. There are particular approaches that in some sense follow the scientific method discussed earlier. The main difference here will be that time series data cannot be laid out in a “designed experiment” fashion. This book goes into much detail about the process, methods, and technology for building these multivariate in X time series models while taking care to find the drivers of the problem at hand.
With regard to process (previously discussed), various authors have reported on the process for data mining transactional data. A paper by Azevedo and Santos (2008) compared the KDD process, SAS Institute's SEMMA (Sample, Explore, Modify, Model, Assess) process and the CRISP data mining process. Rey and Kalos (2005) review the Data Mining and Modeling process used at The Dow Chemical Company. A common theme in all of these processes is that there are many Xs, and therefore some methodology is necessary to reduce the number of Xs provided as input to the particular modeling method of choice. This reduction is often referred to as variable or feature selection. Many researchers have studied and proposed numerous approaches for variable selection on transaction data (Koller 1996, Guyon 2003). One of the main concentrations of this book will be on an evolving area of research in variable selection for time series type data.
At a high level, the data mining process for forecasting starts with understanding the strategic objectives of the business leadership sponsoring the project. This is often secured via a written charter that documents key objectives, scope, ownership, decisions, value, deliverables, timing and costs. Understanding the system under study with the aid of the business subject matter experts provides the proper environment for focusing on and solving the right problem. Determining from here what data helps describe the system previously defined can take some time. In the end, it has been shown that the most time-consuming step in any data mining prediction or forecasting problem is the data processing step where data is defined, extracted, cleaned, harmonized and prepared for modeling. In the case of time series data, there is often a need to harmonize the data to the same time frequency as the forecasting problem at hand. Then there is often a need to treat missing data properly. This may be in the form of forecasting forward, backcasting or simply filling in missing data points with various algorithms. Often the time series database has hundreds if not thousands of hypothesized Xs in it. So, just as in data mining for transactional data, a specific feature or variable selection step is needed. This book will cover the traditional transactional feature selection approaches, adapted to time series data, as well as introduce various new time series specific variable reduction and variable selection approaches. Next, various forms of time series models are developed; but, just
