Forums  > Trading  > How high a Sharpe is considered "good?"  
     
Page 1 of 4Goto to page: [1], 2, 3, 4 Next
Display using:  

Rickets


Total Posts: 5
Joined: Jan 2007
 
Posted: 2007-01-27 21:12
I was just curious as to what kind of Sharpe for a trading system is considered "good." Two? Three? Five? Eight? This would be for live trading results, not a backtested history, and measured using daily returns.

Since there are many variables involved here, let's say we restrict this to systems that:

1. Do not sell options, or attempt to replicate short option positions

2. Do not engage in market making. I'm most interested in evaluating the quality of a directional strategy, not one that collects the spread and employs only minor directional views.

3. Have at least 3 years of live history.

I realize there are many other variables at work here, such as what asset classes are being traded, etc. But just as a general statement, if you approached a bank and said I have a system with a 0.5 Sharpe, they would probably not be too interested. If the Sharpe were 9, that would be a different story. So what kind of Sharpe is considered "good" in the world of quant trading?





FDAXHunter
Founding Member

Total Posts: 8349
Joined: Mar 2004
 
Posted: 2007-01-28 11:11
That depends on a lot more factors, such as scalability, time frame (which are related, but not 1:1), the market (it's easier to make money in an inefficient market) and a bunch of other stuff.

For a single market, you can have long term systems that may hold positions for 12 months or more. For these, a Sharpe Ratio of > 1.0 might already be quite acceptable. For something that flips it's sign 3 times a day a Sharpe Ratio of 2.5 might still not be that great.

As a very rough rule of thumb, single systems with a Sharpe Ratio of above 3 and that scale reasonably well (note added emphasis) are considered very good. I guess that's the answer you are roughly looking for, without further waffling. Now the only thing to answer is what you consider reasonably scalable... and that depends very much on the house.

Note that some people throw around Sharpe Ratios of way above 10. These are always very constrained in the amount of actual profits they can generate. Sure, you can have a Sharpe Ratio of 25, but if you consider that the most you can do is generate a million dollars a year with it, that's not how you build a scalable business. Scalability is never to be underestimated and indeed, is more important to the business as such than pure risk-adjusted performance.

Having said all that, you can't really pick a single simplistic number to classify what constitutes a good investment. There are lots of grey areas that need to be taken into consideration. In my opinion anyway.

The Figs Protocol.

margarita
NP High Priestess

Total Posts: 322
Joined: Oct 2004
 
Posted: 2007-01-30 08:10

FDax, would you please elaborate a bit more on the time frame relation to Sharpe ratio?  Or refer me to some useful reading material?

Thank you,

margarita


Prada issues high-heeled bonds.

Bachelier
Transparent Doppelgänger

Total Posts: 692
Joined: Dec 2005
 
Posted: 2007-01-30 16:46

I'd ditto FDAX on the inefficient markets and scalability front.

I'm in one, and frankly a Sharpe of 1+ is a yawn, and *no*one* notices. Above 2 and you get attention.

Inefficient and small markets can also, hmmm, "subsidize" Sharpe ratios. I had a Sharpe of 10.4 one month (whoo hoo!), simply because I was away from one ja-mumb-O position in my index that had a big fall out. My portfolio moved not a hair's bredth, the index took the full vol hammer, and I looked like the genius that I am for a brief, shinning month. The portfolio fell to earth and ended the year with a 3.34 Sharpe and 11.37% return, all very very good in my space, but certainly not 10.

As FDAX pointed out, scalability is key. My market is closed and worth about 85 billion Euros. I already feel my own size and could def not scale that Sharpe and return at all. If I was doing that Sharpe and return trading currencies though, you probably would not even begin to see me until I was 82 billlion in exposure.

back to shelling peas.....


"it is regrettable that no Dostoevsky lived near him."

FDAXHunter
Founding Member

Total Posts: 8349
Joined: Mar 2004
 
Posted: 2007-01-30 17:44
margarita:FDax, would you please elaborate a bit more on the time frame relation to Sharpe ratio

Sure. For you, anything.

Allright: First thing about the Sharpe Ratio is that it's not independent of the time frame it is calculated on. Why? Because drift (i.e. average return) scales linearly with time whereas volatility of returns scales with the square root of time (assuming the usual: GBM).
So, an annualized Sharpe Ratio is different than a daily Sharpe Ratio (no surprise there to anyone who's ever annualized a volatlity, just wanted to point that out for starters so there's no confusion).

The second thing is the nature of the trading strategy. Take a buy and hold strategy. This will expose you to the natural volatility of the market. There's nothing you can do about it. You're long USD/JPY, you're going to be tracking the volatility of dollar yen over the period.
So, the only way to improve your Sharpe Ratio over the Sharpe Ratio that's inherent in the market is by altering the sign (and/or value) of the position (say in the range from +1 to 0 to -1). The more often you do this, theoretically the higher your Sharpe Ratio can be because you could extract more from the time series without introducing additional volatility (okay there's some shaky assumptions in this, but humor me).

This is one of the aspects of high(er) frequency trading: It allows you to squeeze the same (or higher) return out of a time series with less risk (assuming that you have an edge in the first place) (The second aspect of high frequency trading is to be able to capture microstructural effects).

Not sure that there is any material out there that goes into this. It's kinda obvious if you think about it.
There might be, I'm just not aware of any of the top of my hat.

I've got to jump here. Bye.

The Figs Protocol.

mib


Total Posts: 354
Joined: Aug 2004
 
Posted: 2007-01-30 19:11
actually, many strategies do not have their vol scaling as square root.

although square root scaling holds not just for GBM, it does not hold for non-trivial return autocorrelations. many long-horizon convergence plays have vol scaling slower than square root at under-convergence horizont periods; regime dependent strategies like trend-follwoing and many CTAs scale faster than square root because of regime switches

besides that, scaling power depends a lot on risk-management strategy used by the fund or FoF

Head of Mortality Management, Capital Structure Demolition LLC

Johnny
Founding Member

Total Posts: 4333
Joined: May 2004
 
Posted: 2007-01-30 19:27

To clarify, the assumption needed for square root scaling of vols (st devs) is the same as that needed for adding variances, which is zero correlation between returns. Hence it holds for GBM or any other framework with zero autocorrelation. It also holds for returns that are not independent but which nonetheless have zero correlation.

Anyway, this wasn't FDAX's main point, which clearly still holds and makes sense etc etc.


heartbreakingly simple

zinmaster


Total Posts: 13
Joined: Jan 2007
 
Posted: 2007-02-15 00:04
> I was just curious as to what kind of Sharpe for a trading system is considered "good." Two? Three? Five? Eight? This would be for live trading results, not a backtested history, and measured using daily returns.

Depends a lot on size. If you can run 0.5 at large size and *uncorrelated* to other strategies, you can get a lot of capital from the right sources. There is huge portfolio benefit to uncorrelated strategies. [The real difficulty is because you can't prove your strategy is really 0.5 rather than -0.5...]
Anything over 2 with size is a huge home run. But if you want to trade for, say, Millennium, you probably need a 3 at maybe only $20M capacity. N.B. under your 3 rules, if you think you have a Sharpe of 8, you probably have forward data contamination in your backtest ;-) And if you ran an 8, you are probably doing something equivalent to shorting vol or making markets...

what's it all about, alpha?

DeepQuant


Total Posts: 6
Joined: Feb 2007
 
Posted: 2007-02-16 13:46

That has been my exact experience, high Sharpe strategies tend to have little capacity, bare in mind with a sharpe of 8, you are either looking forward and or you have not accounted for costs. Even RenTec run something like 1.8 and they are in a totally different class to the rest of us mortals.  Capacity is often overlooked by the system builder who is hunting for hi sharpes, yet  from a biz perspective a high SR system is of little to commercial value if you can't make money out of it. North of 1.25 would make you a serious player, take AHL, they manage 10bln on what is claimed to be SR=1.00, but in reality is more like 0.5. furthemore, thier system is almost noddy, in contrast RenTec were closed at 6 yards running at 1.8 coupled with a super sophisticated ensemble. It really depends what you want, if you start of trading small then the high SR will help you, but as you make more profits you become more limited from a reinvestment perspective. If your strategy has an SR of 8, and it is costed, it maybe a very high frequency beast, which means that a slight miss-calc in the costs could render the strategy useless, if anything, increase the costs and see if you are able to retain the bulk of you SR. Strategies that place orders on either side of spread can often result in an underestimation of costs. If after doing all of this you have an SR of 8, you need to get into the market. Personally, I look to the portfolio to deliver the SR, you will find that some 9 parts in 10 come from diversification, and only 10% is delivered through getting a stong predictor. In this respect it is easier to get two uncorrelated strategies of SR=1, yielding a portfolio SR of 1.4, instead of a single  strategy with SR=1.4. The sqrt(N) argument is a powerful one, and not to be underestimated. Go for robustness as opposed to SR, recall all those horror stories of neural nets which worked fantastically well in sample yet failed dismally in live trading... no robustness, over fitted....

 

 


SirAppleby


Total Posts: 182
Joined: Dec 2006
 
Posted: 2007-02-17 18:45


I have found Sharpe ratios to be a starting point but lacking in one key respect. You mentioned that you were looking at directional traders, and most of their strategies are very much path dependent. This is not inherently bad, but the long run profitability of a path dependent strategy is directly correlated to that traders ability to increase and decrease leverage at the appropriate times. The Sharpe does not consider that high volatility in a strong trending market is a desirable characteristic.

In option terms, I want my trader to be a call option with positive gamma and low theta risk. The Sharpe fails to reward the high gamma trader. In fact, I can get an infinitely high Sharpe with a negative gamma strategy (until I blow up).

Does anyone know how to evaluate a trader’s performance with a non-linear benchmarking approach?

Patience is necessary, and one cannot reap immediately where one has sown.

Rickets


Total Posts: 5
Joined: Jan 2007
 
Posted: 2007-02-17 19:15
>The Sharpe does not consider that high volatility in a strong trending market is a desirable characteristic.

This is true, though I'd add that high volatility in a strongly mean-reverting market may be equally undesirable. Perhaps the key is ensuring that your test period is "representative," so that high vol and low vol, trending and reverting are appropriately represented.

By this, I mean that evaluating an S&P futures strategy in the 1997 to 1999 period may be inappropriate (high Sharpe is achieved by simply going long), or evaluating a Peso strategy may be inappropriate during 1995 (either you were on the right side of the devaluation, or not).

FDAXHunter
Founding Member

Total Posts: 8349
Joined: Mar 2004
 
Posted: 2007-02-17 19:48
SirAppleby: Does anyone know how to evaluate a trader’s performance with a non-linear benchmarking approach?

You can just replace the denominator in Sharpe by any risk measure, it doesn't have to be historical volatility of returns. You could use a number of risk measures: VaR, ES, etc.
This then leads a number or Risk Ajusted Performance Measures of which the traditional Sharpe Ratio is only a subset of. RAROC for example is pretty much a term for anything where you devide some rate or return by a risk capital charge.

The Figs Protocol.

aaron


Total Posts: 746
Joined: Mar 2006
 
Posted: 2007-02-18 19:46

There are two different ways to think about a Sharpe ratio, and people tend to confuse them (if they think about it at all).

First, it is a hypothesis test that the strategy's long-term expected return is greater than the risk-free rate. Of course, it makes all kinds of assumptions and is often miscalculated, but let's leave those aside for a moment. Assume I could tell you with precision the probability that a strategy's market-beating track record was the result of random chance. What kind of probability would you look for?

I think a reasonable answer is something like 2% to 15%. Anything below 2% means the opportunity is so obvious it would be exploited away if it were real. That doesn't mean I would refuse to invest in it, but I wouldn't give any extra credit for low probabilities below 2%. Anything above 15% is too likely to be the result of a monkey with a trading screen, and I believe there are a lot of monkeys out there. Again, it doesn't mean I wouldn't invest in it, but I wouldn't invest in it on the basis of projections from its track record.

That means I want a Sharpe above 1, or a good reason why not, and I would give no extra consideration to Sharpe's above 2. This assumes the ratio has been properly scaled over the measurement interval. 1.99 and 199 are the same to me. I put most of my effort in making sure the probability calculation is accurate: l look for non-Normalities (especially the small probability of catastrophic results), nonstationarities, varying risk levels, autocorrelation, selection bias and other things that cause me to distrust the Sharpe ratio.

But all of this only addresses the question of whether the strategy is no better than random. I don't want to invest in a random strategy (especially not a 2%/20% or higher fees), but there are lots of non-random strategies I don't want either. Used this way, a Sharpe ratio is a screen to ignore track records that are too volatile to weigh heavily in the investment decision.

The other use of a Sharpe ratio is as a return on risk-adjusted capital. The idea is the amount of capital necessary to support a strategy is proportional to the standard deviation of return. Again there are many assumptions in that calculation and it is often miscomputed, but the idea is still valuable. Assume that I could calculate precisely the capital required to support a track record, that is, the amount of money that could have been invested in the beginning such that the strategy could have been run on a non-recourse basis if the losses exceeded the invested capital. What kind of return would you look for?

While higher is always better, this is where scalability comes in. If you can raise money at 6%, being able to invest $100 billion at 7% is better than being able to invest $1 at 1,000%. There are simple and inexpensive investment options that seem to give annual Sharpe ratios on the order of 0.2, based on long history and as solid economic theory as we have. If you can double that in simple, inexpensive investments, you have a 5-star public mutual fund, which is a money machine for the management company. Hedge funds and prop trading start getting interested at 0.5, for the simplest, cheapest and largest schemes. Above 1.0 the Sharpe ratio is no longer relevant, just the total amount of money that can be run under the strategy.

If I really believe the denominator is proportional to capital requirement, and the computations have been done correctly, I start getting interested at 0.4 and top out at about 1.0. Used as a capital proxy, strategies can be combined to have Sharpe ratios greater than their individual Sharpe ratios. A 0.4 Sharpe ratio with a zero or negative Beta can be more interesting than a 1.0 with a Beta of 1.


SirAppleby


Total Posts: 182
Joined: Dec 2006
 
Posted: 2007-02-19 17:09
It seems that the time period used for Sharpe analysis is probably the most critical factor in determining whether the number is a robust estimate of trader skill. But I have never seen a new manager’s track record where he starts with weak performance and then turns strong. It is always the reverse! So why can’t I measure his risk adjusted performance independent of time?

For example, if I looked at a FX trader who only traded 10 crosses. He has the discretion to overweight any single cross, but the overall portfolio must be constrained to no more than 5x leverage. I read a paper a few years ago which proposed a risk neutral benchmark approach using lookback options. So in the FX case, simply construct an index portfolio of 10 lookback options for each cross, and then continuously roll each over the period required. Weight them equally on a notional basis, but each option will have different delta and gamma depending on current price.

Could I compare the trader’s track record with the lookback index and say that the trader produced good risk adjusted returns? It’s not the holy grail of benchmarks, but I would love to say to a trader that his 30% return sounds good, but really is sub-par on a risk adjusted basis.

Another use of the index would be to spot a trader who is simply short gamma most of the time, and is one “black swan” away from annihilation.

Patience is necessary, and one cannot reap immediately where one has sown.

prophet
Banned

Total Posts: 149
Joined: Oct 2004
 
Posted: 2007-02-23 12:33
I'm seeing different opinions on what consititutes a "good" Sharpe ratio. Are we all using the same forumula? Who subtracts the risk free return and who doesn't?

The problem with subtracting risk free return is the Sharpe ratio becomes variable with respect to leverage or position size. This is a problem for comparing futures systems that are typically leveraged to a desired risk / max drawdown level.

Jaxx


Total Posts: 217
Joined: Nov 2005
 
Posted: 2007-02-23 13:39

ok - why would you equally notional weight it?


Johnny
Founding Member

Total Posts: 4333
Joined: May 2004
 
Posted: 2007-02-23 14:17

The problem with subtracting risk free return is the Sharpe ratio becomes variable with respect to leverage or position size. This is a problem for comparing futures systems that are typically leveraged to a desired risk / max drawdown level.

It is a problem, but it's relatively minor. To first order approximation (i.e. forgetting about margins and frictional amounts) a futures fund keeps its cash on deposit (T Bills, etc) and its risk in futures contracts. The interest earned on deposit is very similar to the risk free rate deducted when calculating the Sharpe ratio. The Sharpe ratio therefore estimates the return from futures divided by the standard deviation of returns from futures. This number doesn't change if you double or halve the number of futures contracts, i.e. increase or decrease leverage/position size.

Managed futures funds tend to put up about 20% of the value of the fund in margin. Therefore if you want to be really fair you should only deduct 80% of the risk free rate when calculating the Sharpe. But with rates so low, that's not a huge difference.


heartbreakingly simple

FDAXHunter
Founding Member

Total Posts: 8349
Joined: Mar 2004
 
Posted: 2007-02-23 14:26
Johnny: Managed futures funds tend to put up about 20% of the value of the fund in margin.

More likely you will post govies as collateral, so then the risk-free interest rate drops out of the picture entirely again.

The Figs Protocol.

Jaxx


Total Posts: 217
Joined: Nov 2005
 
Posted: 2007-02-23 14:50
or if you don't you'll certainley earn libid either way

Johnny
Founding Member

Total Posts: 4333
Joined: May 2004
 
Posted: 2007-02-23 15:26

Absolutely, even more to my point.


heartbreakingly simple

nodoodahs


Total Posts: 227
Joined: Sep 2007
 
Posted: 2007-09-27 20:56

"This assumes the ratio has been properly scaled over the measurement interval."

So if I were backtesting from year ending 1998 to year ending 2006, I should have nine data points, each point being a non-overlapping year. Is this a correct assumption?

"Above 1.0 the Sharpe ratio is no longer relevant, just the total amount of money that can be run under the strategy."

How would one measure the total amount of money that can be run in a strategy? In particular, I am looking at a domestic (U.S.) stock strategy that returns a given Sharpe when filtered for rotating through 19 stocks monthly with a market cap floor of $100mil. If forced to apply the same strategy to 20 stocks monthly with a floor of $1bil market cap, it degrades the Sharpe.

So expanding that into a general liquidity or scalability formula, if I have X stocks that need to be rotated through Y times a year, with a floor of Z in market cap, and an average float turnover of T times in a year in that stock, how much money can be run through it? Do you usually look at this as a fraction of daily liquidity (i.e. no more than 10% of a day's volume) or as a percentage of market cap?


I haven’t seen a beatin’ like that since somebody stuck a banana in my pants and turned a monkey loose.

ig0r


Total Posts: 164
Joined: Jun 2007
 
Posted: 2007-09-28 02:55
If you're planning on turning over entire positions that are more than a few % of daily traded volume you should seriously look at how that will impact your strategy. I'm not sure how market cap applies here, what matters is actual liquidity.

nodoodahs


Total Posts: 227
Joined: Sep 2007
 
Posted: 2007-09-28 03:13

Thnx ig0r.  Depending on how many is a "few" then market cap would be irrelevant, since most of the active stocks have float turnover of 60-90 trading days or so.  I was trying to relate turnover (T) and cap to generate a liquidity measure.

So how many is a "few %?"  Is it 2%, 5%, 10%?

Also very interested in your take on the Sharpe measurement, is the convention 9 years of data = 9 data points?  The same return stream generates different Sharpe, Alpha, Beta if it is done monthly, or 9x12=108 semi-overlapping years, or 9 non-overlapping years.  It doesn't seem to me that there is any easy conversion from one to another.


I haven’t seen a beatin’ like that since somebody stuck a banana in my pants and turned a monkey loose.

ig0r


Total Posts: 164
Joined: Jun 2007
 
Posted: 2007-09-28 03:31
Depends on the stock and how often you need to turnover positions, I wouldn't want to be coming in and trying to buy (or sell) more than 2% of daily traded volume every day, though. 10% once every 6 months, not such a problem. Try it a few times on the stocks you're looking at - take how much slippage you generate from previous bid-ask midpoint and multiply it by 2 (slightly conservative), use that as your transaction cost. When you go live just keep a close eye on how much worse (or better) you do compared to the testing environment.

I don't see much value in the [sharpe ratio] measure, see: some of the reasons mentioned above. One thing you can look at is 1-year trailing and compare that to how the market traded over those periods (pull up simple historical price, as well as implied vols if you can), this could help you uncover factors that you didn't know you were exposed to previously.

nodoodahs


Total Posts: 227
Joined: Sep 2007
 
Posted: 2007-09-28 03:52

OK, 2% per day is a good rough guide for scalability.

FYI as mentioned in the first post, I am testing with about 20 positions per strategy with a check on conditions once monthly, and turnover varies with strategy but is between 40% and 90% monthly.  I haven't checked to see how much recidivism I get on stocks moved out and moved in later in the year....

I'm not a huge fan of Sharpe, but it does seem to be the "Good Housekeeping Seal of Approval" for some folk.  I will be asked at some point for those stats, and want to make sure that I measure them in the "proper" accepted way.  I prefer a simple CAGR/GSD or AVG/STDEV, a high CAGR, a check on drawdown length & depth, and the percentage of months the strategy exceeds 0% should be high, and the percentage of months the strategy exceeds a stock market benchmark should be statistically higher than 50%.


I haven’t seen a beatin’ like that since somebody stuck a banana in my pants and turned a monkey loose.
Previous Thread :: Next Thread 
Page 1 of 4Goto to page: [1], 2, 3, 4 Next