Neural Network Time Series Forecasting Using Recency Weighting
Brad Morantz (Science Applications International Corporation, USA), Thomas Whalen (Georgia State University, USA) and G. Peter Zhang (Georgia State University, USA)
Copyright: © 2008
In building a decision support system (DSS), an important component is the modeling of each potential alternative action to predict its consequence. Decision makers and automated decision systems (i.e., modelbased DSSs) depend upon quality forecasts to assist in the decision process. The more accurate the forecast, the better the DSS is at helping the decision maker to select the best solution. Forecasting is an important contributor to quality decision making, in both the business world and for engineering problems. Retail stores and wholesale distributors must predict sales in order to know how much inventory to have on hand. Too little can cause lost sales and customer dissatisfaction—If too much is on hand, then other inventory problems can occur (i.e., cash flow, ad valorem tax, etc.). If the goods are perishable, it could most certainly be a financial loss. Items that occur over time, as in the number of cars sold per day, the position of an airplane, or the price of a certain stock are called “time series.” When these values are forecast, the accuracy can vary, depending on the data set and the method. This subject has been greatly discussed in the literature and many methods have been presented. Artificial neural networks (ANN) have been shown to be very effective at prediction. Time series forecasting is based upon the assumption that the underlying causal factors are reflected in the lagged data values. Many times, a complete set of the causal factors either is not known or is not available. Predictions are made based upon the theory that whatever has occurred in the near past will continue into the near future. Time series forecasting uses past values to try and predict the future. A slight modification to this concept is the application of recency. What happened more recently is closer to the current situation than the more distant past. The older data still contain knowledge, it just is not as important (or as correct) as the newest information. Things change, life is dynamic, and what used to be may be no more or may be to a different extent. Modification of the training algorithm of a neural network forecaster to consider recency has been proven on real economic data sets to reduce residual by as much as 50%, thereby creating a more accurate model which would allow for better decision making.
Key Terms in this Chapter
Forecast: Forecast is a calculation or estimate of something future.
Analysis of Variance (anova): ANOVA is a statistical method of partitioning the variance to compare within group to between group variance.
External Validity: External validity is one of the four validities as defined by Cook and Campbell that states whether the conclusion from the experiment can be applied to the world in general.
Time Series: Time series is a sequence of data points in time order, usually taken in uniform intervals.
Objective Function: Objective function is a mathematical function to be optimized.
Mean Absolute Percentage Error (mape): MAPE is a non-scaled error metric.
Recency: Recency is how far in the near past.