Applicability of information criteria for neural network architecture selection

In most of the empirical research on capital markets, stock market indexes are used as proxies for the aggregate market development. In previous work we found that a particular market segment of the Vienna stock exchange might be less efficient than the whole market and hence easier to forecast. Extending the focus of investigation in this paper, we use feedforward networks and linear models to predict the all share index WBI as well as various subindexes covering the highly liquid, semi-liquid, and initial public offering (IPO) market segment. In order to shed some light on network construction principles, we compare different models as selected by hold-out crossvalidation (HCV), Akaike's information criterion (AIC), and Schwartz' information criterion (SIC). The forecasts are subsequently evaluated on the basis of hypothetical trading in the out-of-sample period.