Research at Sofia University >
Faculty of Economics and Business Administration >
Statistics and Econometrics / Статистика и иконометрия >
Please use this identifier to cite or link to this item:
|Title: ||Econometric vs. ARTMAP prediction of economic choice|
|Authors: ||Mengov, George|
|Keywords: ||Multinomial logit model, Logistic regression, ARTMAP neural network, Economic choice, Econometrics, Prediction|
|Issue Date: ||20-Mar-2013|
|Abstract: ||Forecasting economic behaviour is an important problem with practical implications for a number of scientific disciplines, including microeconomics, macroeconomics, marketing, and economic psychology. The ability to predict the economic agent’s choice is a coveted goal for both social scientists and market practitioners. In our time, such studies are conducted with field investigations or laboratory experiments. However, the traditional statistical techniques used to build explanatory models with predictive power are of limited capability and have inherent structural deficiencies. Here we show that an artificial neural network of the ARTMAP family forecasts far better than the state-of-the-art multinomial regression the economic decisions of the participants in a laboratory experiment resembling real markets. We found that when the number of options among which one must choose is four, and hence any systematic predictive success above 25% is valuable, Fuzzy ARTMAP achieved 42.28%, while the most popular logit regression model reached 37.87%. This result demonstrates the greater capability of the neural classifier to utilize correlated input factors, which remain underused by regression analysis. Yet, prediction rates such as the attained here are still very low, and could hardly be raised by more sophisticated statistical techniques, but rather should be improved by incorporating more in-depth psychological knowledge about the decision maker.|
|Appears in Collections:||Statistics and Econometrics / Статистика и иконометрия|
Files in This Item:
There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.