Economic Forecasting Using Artificial Neural Networks

In trying to decide upon a topic for this paper, I initially chose an article concerning the relationship between a person’s income and the amount of money spent on his education. However, when I was doing some unrelated research on artificial neural networks, I stumbled across a fascinating article on the internet entitled “Rule Inference for Financial Prediction Using Recurrent Neural Networks.” Although the article itself was not suitable for this assignment, I was intrigued by the broader idea of using artificial neural networks as a functional economic prediction tool. For that reason I ran a keyword search on Econlit to see if there were any relevant articles in economic journals. To my surprise, the search keywords “neural network” came up with ninety-four hits. Of these ninety-four, only seventy-five were journal articles, and roughly thirty of these seventy-five seemed appropriate for the class. From the six such articles that Olin had, I chose the Trippi and DeSieno article because it was the most straightforward and clear. In addition, the results from the Trippi and DeSieno article were citied in many of the other articles on the subject. Trippi and DeSieno’s goal was to predict the behavior of Standard & Poor’s 500 index using “a machine learning-enhanced trading strategy.” The article briefly describes their methodology and presents the results, which indicate that their best model outperformed a trading strategy based on random trades ninety-nine percent of the time. To predict the market behavior, Trippi and Desieno used an artificial-neural-network-based computer model. This type of model is incredibly useful because it can adapt and “learn.” In other words the artificial network will, even in the absence of a human operator, constantly improve upon its results. As Bass describes them, “[artificial neural networks] are capable of performing complex computations, recognizing patterns, and displaying other forms of artificial intelligence.” In the case of Trippi and DeSieno the network was to find a set of rules that roughly described the behavior of the S&P 500 index. The ‘rules’ are not rules in the sense that they can be deciphered by any person, but instead they are retained within the computer for further computations (i.e. a black box model). Once the system has derived these rules, it will use these patterns to predict future positions of the market. The actual mechanism through which the network processes the data was omitted from the Trippi and DeSieno article, but Ntungo and Boyd gave a fairly clear description of a generic artificial neural network in their comparison of the artificial neural network with linear ARIMA models, which I will try to