Santa Clara University

Unclouding the Crystal Ball

How Statistical Models Can Improve Business Predictions

An airline wants to project how many passengers might fly during the holiday season. A water company wants to be prepared for peak demand in the summer months. A firm that issued stock options needs a sense of when employees will execute them.

These are examples of common business problems with multiple factors and variables that call for decision-making that is often based on analysis of data and known patterns. Increasingly, such business intelligence is gained through artificial intelligence — the use of computer software and statistical models to get the most accurate study of the data and its implications for future business activity.

Manoochehr Ghiassi, Professor of Information Systems at the Leavey School of Business, has played a leading role in academic research in this area, working with colleagues to develop better models and the software to run them.

“What we’ve done is add to existing methods, pushing the field one step forward by offering new algorithms,” he said. “The research creates a new tool that can do things better or offer a new solution where one did not exist before.”

A glimpse of what goes into the research is provided by a paper on forecasting time-series events that appeared in the International Journal of Forecasting in 2005.

Lead author Ghiassi and his colleagues, H. Saidane and D.K. Zimbra looked at time-series problems (which involve cycles of activity such as the airline-passenger question) and developed an algorithm (a procedure for solving a mathematical problem in a finite series of steps) to analyze the data.

Along with the model based on the algorithm, they developed software that allows users to input and test data on a personal computer in a matter of minutes.

What they came up with is known as a dynamic artificial neural network that not only analyzes the data but “trains” itself to make more accurate distinctions as additional data comes in, therefore providing a better pattern on which to base future projections.

After developing the model and refining it, Ghiassi and his colleagues  tested it against several other models on three examples that are widely regarded in the field as benchmarks of forecasting literature. These are instances where a significant amount of data has been collected over a period of time on a subject, allowing researchers to apply a model to the information, see how accurately the model forecasts what actually happened, then compare its performance with that of other models.

DAN-2, the model developed by Ghiassi and his colleagues, significantly outperformed existing models based on two commonly accepted measurements of error and deviation. And, he said, it is a model that has a wide range of practical applications, particularly given its ease of use.

The next step in his research, he said, is looking into developing a model that will work for classifying data better, rather than making forecasts. As businesses collect ever more data, he said, there will be a need for better models to analyze the information.

“Whether you’re Safeway, a medical center, or an energy company,” he said, “you’re probably collecting data about your customers, your business and your procedure. The question then becomes, how you can analyze the data to do your business better. Our model offers a tool to help the decision maker.”

HELPING THE DECISION-MAKERS: Manoochehr Ghiassi has been working on dynamic neural networks — self-correcting statistical models for making business predictions.