On Nonlinear Modeling for the Predictability of Equity Return

Non-stationary time series are commonly found in nancial applications. Added to the complexity are the time-varying nature and non-linearity of accurate models for describing the dynamic behavior of these nancial time series. We extend the techniques of cointegration to handle time-varying, non-linear relationship betw een a time series (\news") and its causally a ected time series. The predictability of daily return, as related to the NASDA Q indexes and to a possible NASDAQ-GEM relationship, is investigated based on a proposed \news" model for dynamic changes. The e ectiveness and robustness of neural netw ork models for handling non-linearity is compared with linear least-squares estimation.