Penn Discourse Treebank style discourse parsing is a composite task of detecting explicit and non-explicit discourse relations, their connective and argument spans, and assigning a sense to these relations. Due to the composite nature of the task, the end-to-end performance is greatly affected by the error propagation. This paper describes the end-to-end discourse parser for English submitted to the CoNLL 2016 Shared Task on Shallow Discourse Parsing with the main focus of the parser being on argument spans and the reduction of global error through model selection. In the end-to-end closed-track evaluation the parser achieves F-measure of 0.2510 outperforming the best system of the previous year.