CMStatistics 2016: Start Registration
View Submission - CMStatistics
B0437
Title: Least angle quantile lasso regression Authors:  Jing Huang - European School of Management Technology (Germany) [presenting]
Abstract: The least angle method is used to select variables in quantile regression. Quantile regression has large applications in understanding tail distribution, and variable selection is helpful to study the regression under limited resources. The least angle method chooses variables by selecting the minimum angle between dependent variables and regressors rather than minimizing the distance, in which case least angle method is believed to have fast speed. Furthermore, we found that the current algorithm for selecting variables in quantile regression is fast in speed, but not sufficient to find the optimal variables set, and an exhaustive method is prohibitive. Therefore, we used an approximation to transform the minimization problem into a simple OLS problem turn the selecting problem into an outlier detecting problem. The we suggest several ways, like least trimmed square and PCA (factor analysis), to improve the algorithm to selection variables at each searching step more efficiently. At last, we conduct a simulation study, which shows PCA method has the best performance among all the algorithms we tested.