Hi fellow Quantopians members and staff,
There something that's been bugging me regarding the contest ranking system for some time.
To illustrate my my point, bear with me this small thought experiment:
Let's say that I submit to the contest two perfectly identical algorithms with one only difference, being that one targets a full leverage using Optimize API MaxGrossExposure of 1.0, while the other targets an exposure of only 0.20, so that the initial capital in the first entry is 10m but only 2m in the second one.
Now, since the two algorithms are identical, one would expect that they'd be ex-aequo, right ?
In fact, there's good cause to suspect that the second algorithm would perform better than the first one. For starters, since Slippage scales with capital, Sharpe ratio and Sortino ratio will be slightly higher with smaller capital. In my tests, this effect is small but not negligible.
Off course the returns of the second algorithm would be 5 times lower (approximately since you can't own share fractions...) with current contest commission policies. But since capital is divided by 5, then volatility, beta, and drawdowns will be also much lower. Stability would be also higher, though I'm not sure to what extent. So out of the 7 statistics that Q computes for algorithm ranking, only one would be lower for the second algorithm, but all the other 6 would be better, slightly for risk adjusted returns, and very highly for pure risk measures.
So even though the two submission are exactly the same algorithm, the second one will beat the fist one by a wide margin either IS or OOS.
Now, say, let's imagine a quant desperate to rank higher in the contest, if he wanted to gain an unfair advantage over some other entry with the same kind of performance/risk statistics, he would simply half his exposure, and see how it goes. And do it again, to satisfaction. I call that half-quantism.
I'm not all knowing, but I have reason to suspect that some of the top contestant use this approach (0.1% vol for 10m equity investment, come on....). This is bad form, because it gives an advantage to contest entries with an unproven ability to scale, over entries with with a proven ability to scale up to 10m. Also, it sets a wrong incentive for people to do the same.
Now, don't mistake me, I'm all in for algorithm that vary active capital with time depending on market dynamics, but i don't think that's what we're dealing with here. Maybe my logic is flawed, in that case please enlighten me. but i stand for it for now.
So I want to suggest this to Q staff:
- Create a qualitative badge named 'Fully invested': It would be granted, if the algorithm is X% invested, Y% of the time.
- Replace beta with jensen alpha: it scales linearly with beta, and it would destroy half-quantism.
- Use the CAPM alpha as a metric.