Each forecasting method has its own strengths and weaknesses. Our expert provides an overview, with particular focus on univariate versus multivariate methods.
A good portion of business intelligence data is about scorecarding -- telling us, our clients and our customers where we are right now. It provides the big picture and the associated drill-downs between "year-ago" versus "current year," or "actuals" versus "planned." It is the latter ingredient, the "planned," that is the forecasting part of BI work. As more systems are automated, knowing what just happened is not good enough. Supply change management, retail provisioning, production scheduling and dozens of other future-oriented business tasks need the viewpoint used by the ancient Greek Oracle -- a gaze into the future.
As anyone who puts a fiver in the office football pool knows, consistently good predictions are downright difficult to do. Yet most forecasts are actually backward-looking and consider only one variable (so called univariate forecasts) as the strongest base for going forward. These univariate models have a descriptive name: they're called "auto-regressive," because they use past levels of a variable -- and that alone -- to predict future values of the variable.
Even using so little to go on, forecasters have been able to conjure up some relatively robust univariate methods to predict future trends over a short time horizon, such as next month's or the next quarter's sales for a particular product. Using methods such as moving averages, exponential smoothing and the Box-Jenkins-based auto-regressive analysis (see the Autobox tutorials in References, below) forecasters can identify important factors like seasonal swings or upward and downward trends. In addition, these univariate methods help describe baseline shifts where demand jumps up or down to a new level. They also describe exceptional outliers where spikes or sudden drops, followed by a recovery in demand, occur due to promotions or special events.
In sum, univariate methods have a number of forecasting advantages:
They require less data and are relatively simple to calculate versus multivariate methods
Using dynamic estimations, they track the changing environment fairly efficiently
They're appropriate for part of the 80:20 rule -- 80 percent of products account for only 20 percent of the demand. Univariate methods fit the bill for that 80 percent of products. However, for the 20 percent of products that make up 80 percent of demand, analysts may want to use more than univariate methods.
They're easy to install as automated methods that dynamically choose a forecasting model with the least error prediction, perform final parameter estimation, and generate a final forecast.
It's this last automated capability that has sparked software development from companies like Just Enough Software, with its automated CPFR (Collaborative Planning, Forecasting and Replenishment) tools, or SAS Institute's new HPF (High Performance Forecasting) system, which generates forecasts for tens of thousands of SKUs to be used for demand fitting, inventory control and replenishment. Another product, Delphus Peer Planner, has a virtual forecasting option. In fact, ERP, SCM, and CRM vendors are adding demand forecasting through partnerships (SAP with SAF AG, J.D. Edwards with SPSS, i2 Technologies with Autobox, etc.) or building it on their own (Manugistics with NetWorks, Oracle ERP with Enterprise Planning and Budgeting).
The Nov. 2001 Harvard Business Review identifies demand planning as the major challenge to retailers and manufacturers -- "getting the right goods to the right places at the right prices and the right time." This is an analytic step above scorecarding, using BI analysis and forecasting to help secure better margins and returns with ever-thinner margins.
Univariate Forecasting's Limitations
To avoid forecasting fiascoes, many automated forecasting vendors are providing forecast verification reports that aggregate the individual demand forecasts. In turn, they allow staff to crosscheck, drill down and even change or refine forecasts. Myopic univariate methods are pretty good forecasters when provided with good data and normal events; but when confronted with anomalous events or exceptional conditions, univariate methods' simplicity becomes a liability. Their focus on the past provides little guidance during changing conditions.
In sum, the limitations of univariate methods are as follows:
They don't provide causal analysis. Univariate methods totally ignore identifying the factors that are critical drivers of the underlying pattern of behavior for the dependent variable.
This lack of causal understanding precludes having an early warning system, other than the judgment of experienced staff.
A lack of causal understanding also means neither "what-if" analysis nor broad system simulation are possible, because no multi-variable interactions are tracked.
As a consequence, strategic or long-term forecasting using univariate forecasts becomes highly risky.
So we have reached the heart of darkness in forecasting. Simple univariate methods do indeed provide very useful forecasts in a short time horizon and under static conditions. But they're very poor predictors during changing market conditions. The other option is the use of multivariate methods.
The Agile ArchiveWhen it comes to managing data, donít look at backup and archiving systems as burdens and cost centers. A well-designed archive can enhance data protection and restores, ease search and e-discovery efforts, and save money by intelligently moving data from expensive primary storage systems.
2014 Analytics, BI, and Information Management SurveyITís tried for years to simplify data analytics and business intelligence efforts. Have visual analysis tools and Hadoop and NoSQL databases helped? Respondents to our 2014 InformationWeek Analytics, Business Intelligence, and Information Management Survey have a mixed outlook.