Notations
I = number of style indices used for classification. We consider four indices: Russell 1000 value index (optimization variable, i = 1), Russell 1000 growth index (optimization variable, i = 2), Russell 2000 value index (optimization variable, i = 3), and Russell 2000 growth index (optimization variable, i = 4);
J = number of scenarios (time periods), j={1,…,J} index of scenarios;
= monthly rate of return of the fund, for which the classification is conducted, under scenario; scenarios are equally probable (in the current case study are the monthly historical returns of the Fidelity Magellan Fund);
= monthly rate of return of i-th style index (i = 1,2,…,I) under scenario, scenarios are equally probable;
= random value having J equally probable scenarios, , i = 1,2,…,I;
= random scenario vector;
= vector of regression coefficients (loading factors);
= vector of auxiliary variables;
= loss function, which is the residual of the regression;
= loss function with auxiliary variables;
where = rescaled Koenker and Basset error function, where is the realization of random value on scenario j, ;
= Rockafellar error function. The statistic of this error function equals to the mixed percentile with coefficients k=1,…,K, (see Rockafellar and Uryasev (2011), Example 10: A Mixed-Quantile-Based Quadrangle);
k=1,…,K = weighting coefficients for the mixed percentile,, ;
= PSG Partial Moment Penalty for Loss function with parameter ;
= PSG Partial Moment Penalty for Gain function with parameter ;
= rescaled Koenker and Basset error applied to the residual of the regression;
= Koenker and Basset error applied to the residual of the regression;
= Rockafellar error function applied to the residual of the regression.
Optimization Problem 1
minimizing Rockafellar error function
(CS.1)
Optimization Problem 2
minimizing Rockafellar error function implemented with Partial Moments
(CS.2)
subject to
linear constraint on additional variables
(CS.3)