No More Hustleporn: Let's Talk About Quant Investing

Tweet by Abraham Thomas

investor, advisor, explorer | exited founder @quandl (NDAQ) | ex hedge fund manager | tweets on data, markets, startups | not all those who wander are lost

1/ Let's talk about quant investing! People tend to label a lot of things "quant", but this muddies the very significant differences between various quant approaches.


2/ An options market-maker, a systematic long-short fund, an HFT platform and a trend/reversal macro trader are all quants, but they do very different things.

3/ Instead of lumping them all into the same bucket, I find it useful to think about various categories of quant investing in terms of their "edge". What advantage does a specific quant strategy rely on?

4/ First is "model edge". If you have a model of the market that is better than others, you'll make money.

5/ Models come in different flavours. There are pricing and expectations models, used for example by derivatives traders to value and hedge complex securities. Traders with more accurate models tend to profit at the expense of traders with less accurate models.

6/ Closely related, there are arbitrage and relative-value models, which compare prices across (theoretically) similar assets. Same stocks on different exchanges, index ETFs vs their components, futures vs underlying, basis trades, geographic and physical arb, and so on.

7/ Then there are persistent-inefficiency models, that identify patterns in prices that can be exploited using a systematic trading strategy. Such patterns "should not" exist in an efficient market, but they do.

8/ Rules-based back-tests, technical analysis, trend-following, mean-reversion, momentum, small-cap bias and currency carry trades are all examples of systematic approaches to capturing such opportunities.

9/ (Genuine opportunities of this kind are rare! Most large, liquid markets are extremely efficient. Trading desks are littered with the memorials of investors who found this out the hard way.)

10/ Finally there are macro models, linking asset prices to real world events: growth, inflation, interest rates, exchange rates, business cycles, investment regimes and so on.

11/ Note that many models combine several of these features: they have some macro parameters to be estimated, some no-arbitrage conditions for calibration, they predict some asset values, and then they show that those assets are persistently mispriced in empirical data.

12/ Interestingly, you don't get paid merely for being right; you have to be "righter" than the market. Your model has to make predictions that differ from what's currently priced in, and you profit only when (if) the market converges to your model.

13/ This introduces a whole set of "meta-model" considerations around risk, path dependency, and distributions. Some of the best quant investors are known for having average models and excellent meta-models. "Meta-model edge", if you will.

14/ Model edge is perhaps the best known type of quant edge. Mathematicians and physicists on Wall St, economics and finance PhDs -- this is what they do: devise, solve, calibrate, back-test and ultimately deploy models of the market.

15/ But "model edge" is far from the only advantage that quants try to exploit. A second, increasingly common type of advantage is "data edge".

16/ If you have data that others don't, or if your data is better than others' data, you'll make money.

17/ Data edge operates across multiple dimensions. Accuracy is one; breadth of coverage is another; length of history is a third; granularity is a fourth; timeliness is a fifth; usability (labels, disambiguation, entity res and ticker maps) is a sixth; and there are others.

18/ The presence (absence) of these attributes makes certain datasets better (worse) than others. And traders making decisions based on better data tend to outperform those using worse data.

19/ Even the best models can't overcome the handicap of bad data -- garbage in, garbage out! -- which is why most large quant funds have dedicated data teams whose only role is identifying, acquiring, verifying, cleaning, structuring and managing data.

20/ It's tedious, expensive, often intensely manual, often non-scalable work, but the rewards can make it worthwhile.

21/ For example, I know of a fund that would systematically scan newspapers from the 19th century, and then OCR them with human-in-the-loop error correction to extract price quotations for everything from steel stocks to sugar.

22/ Thanks to this (substantial) effort, they now have far more historical data on which to test their hypotheses -- and correspondingly higher confidence in their results. In a market where basis point accuracy matters, this is a non-trivial (and long-lasting) advantage.

23/ And then there's the whole explosion of "alternative data". Why rely on backward-looking, aggregated, selective financial statements when you can get a real-time, granular, comprehensive view of a company's operations ...

24/ ... using the plethora of new data sources out there: satellites, social media, consumer transactions, web and app store traffic, footfall, supply chain, and so many more.

25/ Model edge and data edge are both "Red Queen" advantages: you have to run as fast as you can just to stay where you where. Proprietary models and novel data sources inevitably diffuse through the market, and their alpha decays.

26/ So quants have to constantly reinvent themselves to stay ahead of the game: by devising new generations of models, or acquiring new data assets. Firms who can do so repeatably tend to thrive; some call this "process edge".

27/ There's a way to beat the market even if you're using models and data that are widely diffused: through "compute edge".

28/ If you can get to the right answer *faster* than anyone else, you'll make money.

29/ Which explains why many quant funds invest heavily in high-powered computation infrastructure. The median tech stack has come a long way from the HP scientific calculators and overnight pricing runs that used to characterize quant in the 1980s.

30/ This is essential given the size and complexity of the core models that many funds use, and of the datasets that power those models.

31/ Sitting somewhere in between "model edge" and "compute edge" is machine learning. ML algos don't have the economic content of traditional quant models; instead, that economic content is abstracted into the pre-algo step of feature engineering.

32/ Intuitively, feature engineering collapses model insights and data assets into a single set of raw material for downstream ML algos to work on. And to run those algos effectively, you need plenty of compute.

33/ A close cousin to "compute edge" is "execution edge". It's not just about predicting the market and identifying the right trades to do; it's also about having the ability to do those trades optimally.

34/ If you can trade more *efficiently* than anyone else, you'll make money.

35/ Real-world transaction costs are a prime destroyer of "looks-good-on-paper" historical trading strategies. And conversely, if you can materially reduce those costs, a number of otherwise unattainable strategies become economically viable.

36/ Strategies don't exist independently of execution, a mistake many beginner quants make.

37/ In fact if your execution edge is strong enough, you don't even need a deep economic model; you can build an entire franchise out of high-frequency market-making.

38/ (That's another arms race btw -- co-location and bare metal servers and hyper-optimized code -- a whole different set of edges to chase after.)

39/ Execution edge isn't just about faster orders or smarter prices; it can also manifest as privileged access to specific flows, or the ability to trade names and ranges that others cannot.

40/ Desks with genuine execution edge tend to think like market-makers: if you offer the best prices, you'll get more flow. More flow means more information, which in turn means you can offer the best prices. Flywheel unlocked!

41/ This is an illustration of one of the most attractive characteristics of quant edges: they are self- and mutually- reinforcing.

42/ Better data enables more accurate as well as more trustable models. Better models enable more accurate pricing for execution engines to target. Better execution opens up a wider range of tradable strategies.

43/ Better strategies and execution (hopefully) deliver better risk-adjusted returns, attracting more assets which in turn enable greater investment in data, in infrastructure and in talent.

44/ Ultimately what quant funds strive for is "network edge". If you combine models, data, compute, execution, talent and process into a single platform, you can create a positive feedback loop. The more trades you make, the better your outcomes (in theory at least).

45/ You could build a quant fund with just one of these edges, but why would you? The returns to pursuing multiple types of edge are increasing with scale.

46/ That's why the best-known quant funds "do it all" -- and why outside observers often conflate all these skills and edges into a single category. Which is where we started this thread!

47/ The end. If you enjoyed this thread and found it useful, I'd appreciate a re-tweet or a follow. Thanks for reading!