I have a model with no buy rules (except in universe) and a sell rule with rank<x. With 15 stocks it seems better with rank<95. With 50 stocks it is better with rank<90. I suspect that rank<90 might be most robust. What would be your view on that on how to decide which rule is more appropriate?
1 Like
Imo if returns scale nicely with position weight (up to around 20%, beyond portfolio math often breaks due to idiosyncratic risk) and rank tolerance “tightness”, that's a good sign for strategy (or ranking system) robustness. Tells me that the system gets the ordering right. It's natural that dilution of the average factor load (by higher rank tolerance or position number) should reduce return. Of course there is also a lower bound for rank tolerance (due to slippage) and concentration (due to idiosyncratic risk) which you shouldn't cross.
I usually test both dimensions with regard to turnover and then choose the highest Sharpe Option. Especially since I combine several strategies in a book (large chunk of the Robustness question imo shifts to book level in this case).
In a stand-alone strategy, if higher position size and rank tolerance gives you similar or only slightly worse returns, I would consider preferring this one due to scalability and idiosyncratic risk reduction.
My sweetspot turnover is 200-300%. High enough to trust the system (trade sample size, frequent, fresh signals, robustness) low enough to ensure tradability (limited slippage with room for error).
95 and 90 are no magic numbers. It mainly depends on signal volatility. I have pure value systems with Rank<99 sell rules which result in the same turnover as a momentum system with rank<80. Highly dependent on factor combination. So just test if the strategy return scales linearly with position weight and rank tightness. If it does, just target the highest position weight and turnover which fits your sweetspot for your goals (20 stocks, 200-300% turnover is a nice start)
In a good old-school multifactor system with nice ordering without additional sell rules, Rank (or RankPos, also try this one) are just a tool to adjust return/turnover to your sweetspot.
2 Likes