Forums  > Risk Management  > what’s the case for sharpe anyways?  
     
Page 1 of 1
Display using:  

Eulers


Total Posts: 1
Joined: Apr 2020
 
Posted: 2020-04-26 09:06
Using metrics to track performance it has always boggled my mind why sharpe is being used. I mean whats the usecase when return / max DD is the end goal anyways? Using martingale/ higher dimensional momentum scale in/out position sizing, it can go as far to really hurt the “sharpe”. What and why do people use sharpe, whats justified?

Short the pops and long the drops

svisstack


Total Posts: 346
Joined: Feb 2014
 
Posted: 2020-05-08 10:55
I would say that maximizing terminal utility or CAGR is the end goal and the return / max DD is just another approximation like Sharpe?

First Commander of the USS Enterprise

kloc


Total Posts: 33
Joined: May 2017
 
Posted: 2020-05-08 15:33
Well, there are Stirling and Calmar ratios if you prefer CAGR/DD type of risk metrics?

Not as popular as Sharpe, but they're pretty standard...

gamerx


Total Posts: 15
Joined: Sep 2012
 
Posted: 2020-05-10 08:25
There are many problems with the sharpe ratio, but same goes for other metrics.

Guess it is the preferred one simply because everyone else is using it. Makes performance comparable across the board this way.

EspressoLover


Total Posts: 434
Joined: Jan 2015
 
Posted: 2020-05-10 22:06
Optimal Kelly sizing is linearly proportional to Sharpe ratio. Therefore for a risk-neutral investor in an infinite period game without leverage constraints, logarithmic utility scales quadratically with Sharpe regardless of higher moments.

Optimal log-returns are equal to Sharpe^2. For example a strategy of Sharpe 0.5 can achieve 25% annualized log-returns at optimal leverage. Sharpe 1.0 can achieve 100%. Sharpe 2.5 can achieve 625% returns. Sharpe 0.1 can achieve 1%. Etc.

Another way to think about this is to consider the law of large numbers. Consider period to period returns of an asset as an i.i.d. process. (For any reasonably liquid and efficient asset, some sufficiently large period will have sufficiently small auto-correlation between periods.) In the long run, the investors log-wealth converges to a normal random variable. The variance of this "cumulative wealth" variable is linearly dependent on the variance of the assets' return variance. Any higher moments converge to zero.

In other words, for a sufficiently long-term investor without leverage constraints, all risk metrics besides Sharpe become irrelevant.

Good questions outrank easy answers. -Paul Samuelson

wquant


Total Posts: 5
Joined: Nov 2019
 
Posted: 2020-07-08 22:05
I am trying to get a better understanding of Sharpe ratios and what it means for moving forward with a backtested strategy (standard long/short equity returns). Here are a few links to what I've been looking at: 

https://www.twosigma.com/wp-content/uploads/sharpe-tr-1.pdf
https://arxiv.org/pdf/1905.08042.pdf
https://www.davidhbailey.com/dhbpapers/sharpe-frontier.pdf
https://alo.mit.edu/wp-content/uploads/2017/06/The-Statistics-of-Sharpe-Ratios.pdf

In these, there seems to be all manner of different statistical tests to ascertain the significance of a backtest Sharpe ratio. 

My question is, amongst practitioners is there a standard calculation/procedure to determine confidence in their strategy or to determine a required track record length. 

Any guidance is much appreciated. Thanks!

DouglasP


Total Posts: 12
Joined: Jan 2014
 
Posted: 2020-07-08 22:38
Could you expand a bit on how to achieve 25% annual with a Sharpe of 0.5? I would assume that includes a fairly decent risk of losing 99% of your capital?

EspressoLover


Total Posts: 434
Joined: Jan 2015
 
Posted: 2020-07-11 21:00
Well, just to illustrate I grabbed daily return data on SPY going back to inception (1993). Realized annualized Sharpe is 0.48, so it's a pretty good example. Annualized return is 10%, so getting to 25% requires adding 150% leverage.

2.5X SPY has a number of major drawdowns. Nothing as serious as 99% of capital. But you would lose 85% of your capital in the GFC, 80% during the dot-com bust, and 70% to Covid. But keep in mind that those losses are fully recovered by 2012, 2006, and (nearly) summer 2020 respectively. 94% of rolling 5-year periods are positive, and the median return is 242%.

That being said, I don't advocate full Kelly Sizing as a practical manner. In particular, I think most people drastically over-estimate their strategy's Sharpe ratio. Mostly because of a combination of over-fitting, over-confidence, performance decay and the Peso problem. Fractional Kelly, with the fraction somewhere between one fifth and one half depending on an honest and introspective Bayesian assessment, is usually prudent.

Good questions outrank easy answers. -Paul Samuelson

rftx713


Total Posts: 125
Joined: May 2016
 
Posted: 2020-07-12 23:50
EspressoLover - I think I know what you mean by the Peso problem, but can you confirm?

doomanx


Total Posts: 77
Joined: Jul 2018
 
Posted: 2020-07-13 11:09
Peso problem refers to the pricing (or in the context EL is using it, the inability to price) rare, significant shifts in the price and/or fundamentals of an asset. As an example, the name comes from the de-pegging of the Mexican peso from the USD. The Peso was pegged to the dolar, but there was a big gap between the IR on Mexican and US bank deposits with the argument that the gap represented uncertainty in whether the peso would be devalued or not. The peg was removed and the Peso tanked.

In stats terms you are trying to estimate a non-stationary process with a small sample size under fat tails. Historical Sharpe doesn't reflect the uncertainty in when the next structural shift might be. As we're in fat tail land you also can't predict when they are going to happen or how severe they might be. Shrinking your Kelly fraction will reduce drawdown in events like that (and it's always better to underbet than overbet if you're not full Kelly, mathematically and practically, see filthys book).

did you use VWAP or triple-reinforced GAN execution?
Previous Thread :: Next Thread 
Page 1 of 1