Seminar Details

Seminar Details


May 10

3:30 pm

Minimizing Downside Risk via Stochastic Dynamic Programming

John Moody


Oregon Graduate Institute of Science and Technology - Computational Finance Program

We propose new techniques to optimize trading systems with respect to downside risk. These techniques are based on recursive algorithms for stochastic dynamic programming. The downside risk criteria include information ratios that generalize the Sharpe ratio be incorporating risk measures based on the Downside Deviation. These measures of risk, also known as the Lower Partial Moments, can be easily used to model a variety of behaviors from risk-avoiding to risk-seeking, by adjusting the desirability of positive or negative skewness in the returns distribution. In a previous paper (1), we proposed a recursive stochastic dynamic programming method called recurrent reinforcement learning (RRL) for maximizing Sharpe ratios and the differential Sharpe ratio. Here, we extend that approach to enable the direct optimization of trading systems with respect to downside risk measures. Extensive empirical results on both artificial and real price series demonstrate the efficacy of the methods. We find that optimizing the Downside Deviation Ratio (DDR) results in trading systems that are better able to avoid large adverse moves than traders optimized with respect to the Sharpe Ratio (SR). This effect is particularly pronounced when the market returns distributions are skewed. The DDR traders take neutral positions more frequently than the SR traders, and achieve smaller maximum drawdowns and higher Sterling ratios (average return over maximum draw-down). These improvements are significant, both statistically and economically.

Joint work with Matthew Saffell:

E-mails -,


(1) Performance Functions and Reinforcement Learning for Trading Systems and Portfolios. John Moody, Lizhong Wu, Yuansong Liao & Matthew Saffell. Journal of Forecasting, vol 17, pp. 441-470, 1998.