PHOTO: Richard Drew/AP Photo

30 April 2008—Taiwanese computer scientists have developed a genetic algorithm—one that evolves to improve its performance—that can predict the impending demise or distress of publicly traded companies. Its creator say it outdoes commonly used financial algorithms at picking the probability of troubles from bankruptcy, check bouncing, takeovers, bank runs, negative net book value, and other financial woes. In tests on publicly traded Taiwanese companies, it spotted flailing firms with 90 percent accuracy two years before they flamed out.

Carlos A. Coello Coello, an expert on such algorithms with the Research and Advanced Studies Centre of the National Polytechnic Institute, in Mexico City, who was not involved in the research, says that for prediction accuracy, 90 percent is “to say the least, remarkable.”

According to the software's coinventor Ping-Chen Lin, an associate professor at the Department of Finance and Institute of Finance and Information at the National Kaohsiung University of Applied Sciences, in Taiwan, the secret to the predictive system's success is the combination of the genetic algorithm's ability to evolve to weigh the value of a set of variables, and the mixing of three types of prediction procedures.

For the period from 1993 to 2004, Lin and her research partner, the late Jiah-Shing Chen, who was a professor in the Department of Information Management at Taiwan's National Central University, in Jhongli City, retrieved the pertinent statistics of 537 domestic companies listed either in the Taiwan Stock Exchange (TSE) or GreTai Securities Market. They then used three-fourths of the companies to train the program and applied the results to the other fourth.

The company statistics used in the algorithm consisted of 39 variables—including net sales, total assets, market value of equity, sales-growth rate, return on total assets, current ratio, and leverage—indicating the financial situation of companies. Lin and Chen then used a genetic algorithm to obtain different weights for each according to how well they correlated with the companies' final financial situations.

“Imagine [how] a chromosome having 39 genes evolves,” says Lin. “Natural evolution has a tendency to optimize the situation. Its application in financial analysis can reduce irrelevant variables.”

Once the genetic algorithm has selected and weighed the variables, they are fed into a second stage of the program, a hybrid prediction program. It integrates the advantages of three different prediction methods: discriminate analysis, logistic regression, and neural networks. The first two are traditional statistics methods, explains Lin. The last is based on an artificial neural network—an interconnected collection of software constructs that mimic some of the properties of brain cells.

“The results show that the combination of three models leads to the highest accuracy rate,” says Lin. Indeed, the mean forecasting accuracy of Lin's predicting system for eight financial quarters prior to a company's failure is roughly 90 percent.

Lin says this hybrid approach could help fund managers, stock investors, banks, and others avoid risk. Taking the loan departments of banks in Taiwan as an example, Lin says that a final decision to lend money to a company is sometimes made based on past experience and the results of existing programming software tools. Those tools are designed using traditional statistical methods and cover certain indicators, which might be less suitable than those selected by the genetic algorithm, she says.

The algorithm was developed and tested using only Taiwanese companies, but a similar predicting system can be easily constructed for companies in other countries “as long as researchers select variables appropriately and have enough target companies to conduct experiments,” Lin says.

Lin reported the results of the system in the March 2008 issue of the International Journal of Electronic Finance. The same month, she published a Chinese-language book (in Taiwan) further detailing the research. Now Lin's team is working on improving the prediction system in a bid to forecast whether a company will default on a loan. She says banks can adjust the terms and conditions based on the forecast to minimize the risk induced by nonperforming loans.