Time Delay Neural Network for Small Time Series Data Sets

For large data sets neural networks have better time series prediction performance than conventional methods 6. However, when neural networks are trained with meager data, the problem of overfitting seriously decreases their performance. The objective of this work is set apart from previous neural network approaches by focusing on noisy data of limited record length. Observation of monthly slaughter hog prices from January, 1965 to July, 1987 are used. The work uses two techniques to overcome the problem of overfitting: the validation method and the use of time delay neural networks (TDNN).