Risk-Adjusted On-line Portfolio Selection

The objective of on-line portfolio selection is to design provably good algorithms with respect to some on-line or offline benchmark. Existing algorithms do not consider 'trading risk'. We present a novel risk-adjusted portfolio selection algorithm (RAPS). RAPS incorporates the 'trading risk' in terms of the maximum possible loss. We show that RAPS performs provably 'as well as' the Universal Portfolio (UP) [4] in the worst-case. We empirically evaluate RAPS on historical NYSE data. Results show that RAPS is able to beat BCRP as well as several 'follow-the-winner' algorithms from the literature, including UP. We conclude that RAPS outperforms in case the assets in the portfolio follow a positive trend.