Statistical arbitrage dynamics is driven by a stationary, autoregressive process known as mispricing. This process approximates the value in time of a portfolio weighted equally to the elements of a cointegration vector of the log-prices processes of related instruments. Statistical arbitrage involves taking either long or short positions on a portfolio according to predictions of mispricing. This paper offers a theoretical analysis of cointegration testing under the conditional heteroscedasticity of the innovations process. Cointegration testing is used in the procedure of searching for the log-price processes of the related instruments that will form a statistical arbitrage portfolio. We also investigate dynamic characteristics of the mispricing process, which is a linear combination (cointegration vector elements are coefficients of it) of related log--prices processes for which the (T)VECM-MGARCH model class is assumed. Under this model assumptions making precise predictions on mispricing process based on past realizations are difficult. This paper can be treated as a starting point for an empirical analysis of statistical arbitrage portfolio construction. Reference is made to theory to describe the challenges which can be faced in constructing a statistical arbitrage portfolio based on cointegration, in modelling the dynamics of mispricing, and in prediction where the innovation process is conditionally heteroscedastic.