Genetic Algorithm bull bear rsi strat tuning - how to measure strategy effectiveness?

[coiner2018]

I've written a simple .NET Core app which is using Genetic Algorithms to tune the RSI BULL BEAR ADX strategy.

What I'm trying to work out, is, when measuring the "fitness" of a given set of strategy parameters, is it best to pick the parameter values that simply lead to the highest returns or the highest Sharpe Ratio?

I started off using the Sharpe Ratio - as I assumed this might lead me to tuning the strategy to be a mix of "reasonable risk" and "reasonable return" but I've found that in a lot cases, the highest Sharpe Ratio numbers can often produce pretty poor returns (whereas lower ratios can produce very good returns).

Anyone else done some work/research on this?
 
(PS for anyone interested, I'm using the GeneticSharp Nuget package for the GA implementation)

[Gryphon]

Lots of factors to consider here... For simple metrics, look at the % profitable trades, average positive and negative trade values, and exposure time.

A strategy that makes thousands of small trades with very small exposure times (couple of minutes) will not work in live as you are very unlikely to be able to fill the orders quickly enough. GA on RSI strategies will very often lean towards this style of trading in backtesting if you are targeting profit alone - and report multiple millions of % profit.

[coiner2018]

Lots of factors to consider here... For simple metrics, look at the % profitable trades, average positive and negative trade values, and exposure time.

A strategy that makes thousands of small trades with very small exposure times (couple of minutes) will not work in live as you are very unlikely to be able to fill the orders quickly enough. GA on RSI strategies will very often lean towards this style of trading in backtesting if you are targeting profit alone - and report multiple millions of % profit.

Thanks - useful points re. exposure time and the number of small trades - I hadn't considered those.