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, Hosmer-Lemeshow 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.