Should one follow a purely Quant approach, as seems increasingly popular today, or should one on the contrary combine quantitative and qualitative ideas for best results in trading? 

Intuitively mixing qualitative judgment with quantitative signals matches pension funds' desire to blame someone if something goes wrong, so intuitively it should command higher fees and more assets. Less cynically qualitative judgment is harder to replicate. Theoretically. In reality I find most people's qualitative judgment is just a randomly executed quant system.

For similar reasons I can imagine purely quantitative processes performing better, when the sole mandate of the manager was to define methodologies to turn systems on then subsequently turn them off. But it's hard to ignore the effect of AQR on fees and industry events like Cohen plowing into Quantopian, as both worsening pricing and increasing competition in the quant space.

I'm trying to figure out what method is the best to pursue. Should I be reading the earnings transcripts, talking to management, using the software companies make and ad platforms of tech companies, doing my best to make a robust qualitative view? Or should I be improving my use of machine learning models and getting more proprietary data sets?

More simply, does the next 20 years in have asset management have a stronger bid for the qualitative, the quantitative or the hybrid?

I would be most grateful for your wisdom.

Bill Rafter writes: 

Let's say you have a quant "system" that you have tested and it has a positive expected value that is of interest. Adding some qualitative/anecdotal tinkering on top of your tested program has a real risk of lowering your expected value (assuming you have no ability to test your tinkering.) So why tinker? Well, it's human nature to do so, and by tinkering you might find something better. Okay, then put 90 percent of the capital into the program with the tested positive expected value and experiment with 10 percent, or just hold that latter capital back for when you positively test another system.

BTW you might want to read Ralph's thoughts on how much to bet.

The tougher part is coming up with the "system". Obviously test everything, especially your assumptions. From reading your note I see that you might have some untested assumptions. For example do you think earnings are important, something which I myself do not know? I'm not saying they are unimportant, just that I don't know. For example we do a lot of macroeconomic forecasting, but we never trade based on it because we have learned that the market does what it wants to do, and not necessarily what the economic numbers suggest. And also we know that a lot of the macro releases are fudged.

One thing you should give serious consideration to is which time venue you will target. Unless you have the right infrastructure it will not be high frequency trading. So will it be days, weeks, or much longer? That will dictate the type of approach you pursue and your research. If it will be very long term, then you have to get deep into company research.

The people who care about earnings tend to look at the much longer time frame. Meaning that your capital is exposed for a long time during which lots of randomness can work their evil ways. [The factors that we are most capable of dealing with are momentum and sentiment, and consequently our time frame of interest is shorter, say 4 days to 6 months.] So identify your strengths and go with them, particularly if those strengths differ from that of the crowd. If you don't know what your strengths are, be prepared to put in a lot of time on research. Minimize your trading during that period otherwise you will not have seed capital to trade when you acquire the skills. You know that, but it bears repeating.

Be prepared for the counterintuitive. For example, when we first acquired the computer skills to do the research we did "test 1". Test 1 was "if you know the market is going to go up, which stocks do you buy?" We assumed it would be the high beta stocks, as they would go up more. But they didn't. Turns out that beta is backward-looking and going forward the high-beta moniker just means higher volatility, which is a negative. So test everything and assume nothing.



Hey guys,

So after watching the old PTJ video a long time ago I used to have a strategy that would allocate to things that had a tendency to go up after large factor moves. The Chair also mentioned strategies like this in The Education of the Speculator.

For example, Priceline would tend to underperform its peers on a t+2 basis so long as the Euro was in free-fall. So the strategy would keep trading Priceline versus its peers based on movements in the Euro. The qualitative thesis sounds great - the street is slow to adjust its expectations relative to the movement of real time macro assets that have material impact on their forward earnings.

Except for one thing.

In 2011, this strategy started falling apart and I stopped using it. From 2012-2014 it has gone into full free-fall. Moves have become coincident, basically, and over-react. So once EUR starts puking Priceline will immediately start gapping versus its peers, lower. On a t+2 basis it will tend to correct this over-response. I tried to improve the strategy to be more clever (like isolate factor exposure by stripping out fixed income from yen exposure, because things that move up with usdjpy tend to also move up with usgg10yr). It didn't matter/ reverse the breakdown.

My guess is that this is due to the increasing prevalence of quant shops who can run sophisticated lagging correlation analysis minute by minute. I've noticed that people on the list ascribe a lot to backtests like this. Do you have any qualitative analysis of your returns these days versus your returns before 2011 that you'd be willing to share? Is it harder to make $ now?

My gut instinct in response to the BOJ is to pile into things with Japan QE factor exposure (such as thing with a high weight in EWJ, or yield assets in Brazil - Japan FDI in Brazil has been in a structural uptrend for quite some time). But supporting the idea that an overreaction has already occurred, some Brazil yield assets moved up as much as 10-12% on Friday after being run up previously (mysteriously haha, obviously someone knew what was about to happen in Japan). But if we're in an existential market where this is already priced in - the trade is to fade it, not pile in. It'd be interesting to hear your thoughts on the subject.





 I have been thinking about what could be a good set of criteria to measure trading (strategy) performance for individual traders.

The criterion of average return divided by the variance of the returns seems to have its shortcomings. One reason is that some large positive returns can cause the variance to go up resulting in an indication by the criterion that the performance deteriorates. But some large positive returns are good to have.

Other criteria like Sharpe ratio seem more suitable for institutions.

I think using properties of the linear regression line of the cumulative return curve might be a better choice.

Two useful properties are the slope and the "width" of the linear regression line. By "width" I mean the deviation of the cumulative return curve around the linear regression line.

A good performance should have high slope on the one hand. And if we do not consider reinvesting profits, it should have narrow "width" around the linear line.

So then the value of slope/width seems meaningful.

If we take the linear regression line as a risk free benchmark, then this value may be very similar to the definition of Sharpe ratio, but practical for individuals.

Would anyone please comment on the pros and cons of this, or any other better ways to measure performance.

Alexander Good writes: 

Great post!

I think it makes sense to measure linearity of PNL and convexity separately so I agree with you that R sq is a good one to employ. I am curious how width differs from the strategy's std though…

One thing that you can do as a cheap proxy is median return * sqrt(252)/std return and then for skew then have a (rolling max peak to trough draw down)/(rolling max peak to trough draw up).

You can benchmark your strategy vs. bonds, the S&P and a traditional 60-40 mix or your other strategies. It's very hard to beat a vol weighted portfolio of stocks and bonds so it's a good benchmark in my humble opinion assuming you're trading your PA and you don't have large retirement holdings. I assign different weights to skew and median return depending on my portfolio construction.

In portfolio construction you'll often find things with strongly positive skew have good inverse correlation to market PNL series and are typically 'long vol' (idea ripped off AQR's value and momentum everywhere).

Trending strategies frequently have very positive skew (momentum) whereas mean reversion tend to have skew that looks like the S&P (value). So if I'm net long beta my marginal utility of doing trending models is higher whereas if I'm net short I tend to size up mean reversion strategies.

Would be curious to know what other people are using/ how other people think about this/ if they have good papers on the subject. 

Leo Jia writes: 

Aren't they different?

std of returns has this term: (Ri - mu)^2, where mu is the same for all i's.

The width has this term instead: (CRi - Vi)^2 where Vi is the value on the linear regression line at time i and is all different across all i's.

Alex Castaldo writes: 

Personally I just like to look at the equity curve visually, and it is not difficult to store large numbers of graphic files in a folder and quickly "flip" through them by hitting a key on the computer.

But for automated evaluation Leo's two criteria (slope of regression, and "width around the regression" (which is also called the SEE or standard error of regression textbooks) make sense to me.

However I know there are many other criteria that have been proposed. There is one with a foreign name that I think starts with "v" but that I can't remember. I am sure some people here know what I am talking about, it was much blogged about 2 or 3 years ago.

In looking for it I accidentally googled another measure of equity quality, the k-ratio , that believe it or not has 3 different versions.

Any other ways to measure equity curve "quality"?

anonymous writes: 

As with many things involving non linear information, my experience suggests that one must mix, blend or combine different 'quantities' to form a unique and proprietary time series.

For example, some form of 3D 'curve' that combined the three quantities return, AUM & volatility that gets thicker as AUM in the strategy grows and changes colour as volatility of returns increases perhaps… 

Ralph Vince writes: 

percent of 6 month periods underwater
percent of 1 year periods underwater
percent of 2 year periods underwater

percent of time at equity highs
percent of time within 1% of equity highs
percent of time within 5% of equity highs
percent of time within 10% of equity highs
percent of time within 20% of equity highs

I have all of these programmed up in javascript which you can peruse at and click the "compare" tab. 


Resources & Links