William Remus

Artificial neural network models for forecasting and decision making

Abstract Some authors advocate artificial neural networks as a replacement for statistical forecasting and decision models; other authors are concerned that artificial neural networks might be oversold or just a fad. In this paper we review the literature comparing artificial neural networks and statistical models, particularly in regression-based forecasting, time series forecasting, and decision making. Our intention is to give a balanced assessment of the potential of artificial neural networks for forecasting and decision making models. We survey the literature and summarize several studies we have performed. Overall, the empirical studies find artificial neural networks comparable with their statistical counterparts. We note the need to consider the many mathematical proofs underlying artificial neural networks to determine the best conditions for their use in forecasting and decision making.

Graduate students as surrogates for managers in experiments on business decision making

Abstract Most laboratory research in business has used graduate or undergraduate business students as subjects. These subjects have been assumed to be suitable surrogates for business managers and, therefore, the experimental results should be applicable to real businesses. This experiment compared the decision making of line managers no lower than second level with that of students with no managerial experience. The level of educational achievement was held constant between the two groups. There were no significant differences between these two groups in making production scheduling decisions.

Time series forecasting using neural networks: should the data be deseasonalized first?

This research investigates whether prior statistical deseasonalization of data is necessary to produce more accurate neural network forecasts. Neural networks trained with deseasonalized data from Hill et al. (1996) were compared with neural networks estimated without prior deseasonalization. Both sets of neural networks produced forecasts for the 68 monthly time series from the M-competition (Makridakis et al., 1982). Results indicate that when there was seasonality in the time series, forecasts from neural networks estimated on deseasonalized data were significantly more accurate than the forecasts produced by neural networks that were estimated using data which were not deseasonalized. The mixed results from past studies may be due to inconsistent handling of seasonality. Our findings give guidance to both practitioners and researchers developing neural networks. Copyright

Judgemental forecasting in times of change

Abstract This paper reports a study which examines the ability of people and statistical models to forecast time series which contain major discontinuities. It has often been suggested that human judgement will be superior when circumstances change dramatically and statistical models are no longer relevant. Using ten time series that contained five different discontinuities and two levels of randomness, the results indicated that people performed significantly worse than (parsimonious) statistical models. This occurred for the segments of the time series where the discontinuity was to be found and for the subsequent segment where the series was stable. People seemed to change their forecasts in response to random fluctuations in the time series, identifying a signal where it did not exist. This was especially true for the series with high variability. The implications of these results for forecasting practices are discussed.

Neural Networks for Time-Series Forecasting

Neural networks perform best when used for (1) monthly and quarterly time series, (2) discontinuous series, and (3) forecasts that are several periods out on the forecast horizon. Neural networks require the same good practices associated with developing traditional forecasting models, plus they introduce new complexities. We recommend cleaning data (including handling outliers), scaling and deseasonalizing the data, building plausible neural network models, pruning the neural networks, avoiding overfitting, and good implementation strategies.

Neural network models for intelligent support of managerial decision making

Abstract Neural networks can provide advantages over conventional models of managerial decision making including being easy to embed in intelligent systems and learning from the data presented rather than requiring human interaction. This article reports a study of the ways in which neural networks can be used to model managerial judgment. In this research, we built composite neural networks and compared their performance model closest in philosophy to the best classical composite model gave the best economic performance.

Neural network models as an alternative to regression

Neural networks can provide several advantages over conventional regression models. They are claimed to possess the property to learn from a set of data without the need for a full specification of the decision model; they are believed to automatically provide any needed data transformations. They are also claimed to be able to see through noise and distortion. An empirical study evaluating the performance of neural network models on data generated from three known regression models is presented. The results of this study indicate that neural network models perform best under conditions of high noise and low sample size. With less noise or larger sample sizes, they become less competitive. However, in two of the three cases, the neural network models were able to maintain mean absolute percentage errors (MAPE) within 2% of those of the true model.<<ETX>>

The impact of presentation media on decision making: does multimedia improve the effectiveness of feedback?

This paper reports an experiment that examines the impact of presentation media on the effectiveness of feedback information in a decision-making task. The study was based on control theory and resource-matching theory (RMT). A laboratory experiment with 72 participants was conducted in the context of providing feedback when using a decision support system. Consistent with hypotheses derived from theory, when negative feedback is delivered using non-vivid (textual (TEXT)) messages, it induces a higher level of subsequent task performance than when it is delivered using vivid (multimedia (MM)) messages. On the other hand, contrary to the expectation from Control theory, negative feedback, in general, does not lead to a higher level of subsequent task performance. The implications of the findings on multimedia are discussed.

Does reliable information improve the accuracy of judgmental forecasts

Abstract This study investigates peoples ability to use information when forecasting time series. Previous studies of time series forecasting have emphasized the importance of additional information in improving the accuracy of the final forecasts. In this study, the subjects were presented with perfect and imperfect information about the future direction of the time series; also there was a control condition with no additional information provided. Results indicate that the more accurate the information provided, the more improved the quality of the forecasts. The subjects, however, were not able to fully use the information and thus underperformed simple statistical forecasting models.

Evidence and principles of functional and dysfunctional DSS

Though the primary objective of DSS is to improve decision making effectiveness, recent DSS experiments find that this objective is accomplished in disappointingly few cases. These DSS experiments are reviewed and their results are analyzed using several general empirical findings in psychology and management. These findings include the dubious robustness of normative decision aids, the sensitivity and variability of decision making behavior, and the realities of organizational decision making processes. Finally, the DSS research mission is discussed.