(A) study on arbitrage pricing theory by artificial neural networks

Abstract

Arbitrage Pricing Theory (APT) is the main research field of finance. This thesis is willing to lighten whether the macro economic variables used instead of factor analysis are priced. And to overcome the limitation of traditional methodology-the time-invariant assumption of factor loading, we used a new technique called artificial neural network. With the technique factor selection and factor loading estimation method are developed. Macro-economic variables selected as common factors are inflation, $M_2$, house price index, industry production index, oil price, exchange rate, bond risk premium, and market reture. 26 industry indices are used as dependent variables. We have trained the neural network from July, 1982 to December, 1988. Bond risk premium, market return, house price index, and $M_2$ are shown to be significant on returns of stocks over the training period. At each time bond risk premium and market return are shown to be almost significant. But $M_2$ and inflation alternate signficance and insignificane. House price have the great effect on the returns at a certain period. We can say from empirical results that there is 3 or 4 commom factors in Korean stock market and factor loading and factor risk premium are time varying.