Jul 28, 2017 · 1 min read
This is good question. In theory it is not true but in practice and for applications in the specific domain of trading strategies I have found out that SVM and BLR produce close results. This is based on skin-in-the-game with hundreds of hours spent working on data from competitions. Even with a large number of features I have found out no advantage in using kernel SVM over BLR unless one fiddles with parameter long enough. For other domains, especially where features have spatial rather than temporal characteristics, the difference may be large.
