

Explaining Danica v0.5 optimization choices  Volatility effect on the position of the optimum 
Written by Forex Automaton  
Thursday, 14 January 2010 13:45  
Page 2 of 2
The middle and halfwidth of the bands around the data points in Fig.2 represent the mean and standard deviation of the set of the individual Pearson correlation coefficients for the 6 major forex exchange rates, treated as independent measurements.
Fig.3 is an attempt to quantify the impact of various volatility regimes into the overall performance figure. The colored data sets are obtained by applying symmetric cuts on the forecastreality product as discussed above, the one of the type abs(product) < cut and calculating the Pearson correlation coefficient for the data that satisfies the cut. Because with increasing volatility, both the actual logarithmic returns (by definition) and the forecast ones increase, such a cut effectively limits volatility from above. The positive net performance is seen to be largely created by high volatilities. The study reveals that the higher Fred values (starting with perhaps 15) split the different volatility regimes in terms of performance. This is undesirable since low volatility regimes is where the market spends most of its time. Indeed there is no assurance that the recent extreme volatility will repeat itself. This reminds me of the conclusions from the full trading system optimization (including the tradeidea selection and portfolio balancing components disabled in Danica) attempted in summer 2009: in the absence of Pearson data like Fig.2, I was led (by return/risk considerations) into the upper portion of Fred, to discover that the best "high Fred" systems traded almost exclusively during the peak of the financial panic.
While choosing the Fred position (the choice for v0.5 of the trading system is indicated by the green arrow in Fig.4) I was balancing the goal of having high "no cut" Pearson for "close", the goal of having high Pearson for "high" and "low" (there, product cuts like Fig.3 and 4 make little difference to the position of the maximum), and the goal of not getting in the area of too much variety in the behaviour of the colored data sets in Fig.3 and 4. What is currently believed to be a reasonable compromise is indicated by green arrow in Fig.4. 

Last Updated ( Wednesday, 14 April 2010 12:45 ) 