The goal of this study is to automate the Autoregressive Moving Average(ARMA) model building and forecasting procedures. To achieve this goal, an Artificial Neural Network(ANN) approach is adopted. Thus, the key research issues are:
1. Which statistic is the most effective as a feature extractor for ANN?
2. How to design the ANN?
3. What is the performance of ANN approach?
4. Can the ANN replace not only the identification step, but also the parameter estimation step?
To proceed with the issues, we decompose the research into the following three projects.
1. When we use the ACF and PACF as statistical features, what is the performance of ANN approach for ARMA model identification?
2. When we use the Extended Sample Autocorrelation Function (ESACF) as a statistical feature, what is the performance of ANN approach for ARMA model identification? What is the performance of this approach in comparison with that of the first approach?
3. When a set of time series data is fed directly into the ANN, what is the comparative forecasting performance in comparison with the above two approaches?
The purpose of the first and second projects are to replace the role of human expert in ARMA model identification step by the ANN, while the purpose of the project is to replace the whole modeling procedure by an ANN which acts as a system modeling tool in the time series generating process.
In the first project, two independent Multi-Layered Perceptrons (MLP``s) with two hidden layers are designed as pattern classification networks(PCN) which identify the order p and q from the ACF and PACF patterns, respectively. Maximal region covering strategy is used to prepare the training data sets and to increase the information values of the ACF and PACF. the learned MLP``s are tested with both artificially generated and real world time series data sets. The ANN approach is proved to be useful in automating the ARMA model identification step.
In the second project, since the performance of ...