One-day prediction of state of turbulence for financial instrument based on models for binary dependent variable

This paper proposes an approach to predict states (states of tranquillity and turbulence) for a financial instrument in a one-day horizon. The prediction is made using 3 different models for a binary variable (LOGIT, PROBIT, CLOGLOG), 4 definitions of a dependent variable (1%, 5%, 10%, 20% of worst realization of returns), 3 sets of independent variables (untransformed data, PCA analysis and factor analysis). Additionally an optimal cut-off point analysis is performed. The evaluation of the models was based on the LR test, HosmerLemeshow test, GINI coefficient analysis and KROC criterion based on the ROC curve. Nine combinations of assumptions have been chosen as appropriate (any model for a binary variable, the dependent variable defined as 1%, 5% or 10% of worst realization of returns, untransformed data, 1%, 5% or 10% cut-off point respectively). Models built on these assumptions meet all the formal requirements and have a high predictive and discriminant ability