Interview with Tammer Kamel
Tammer Kamel is the founder and CEO of Quandl - a data platform that makes financial and economic data available through easy-to-use APIs.
Listen to this two-part interview with him.
Note: The interview is audio-only; closed captioning is available (CC button in the player).
===
And the idea here was, it was a smart idea, and I can't take credit for it entirely.
But the idea here was if you mix normal distributions and drew from them at different probabilities. So for example, you had a distribution with volatility of sigma.
> So you might have a distribution of returns for one strategy or one stock.
And so you're talking about the distributions for maybe different strategies or different stocks and then potentially combining them in some way.
And everybody assumes those distributions are Gaussian, right? So the bell curve-
Right. the problem is that reality doesn't intersect with them that much, right?
But one nice thing you can do is you can actually simulate reality a little bit better if you use not one, but two Gaussian distributions. And as your random variable moves, you draw usually from the first normal distribution. But then some random probability, you draw from a distribution with a much higher standard deviation.
Thus simulating these rare events and creates the fatter tails, and is great for risk management. It's predicted power is minimal, because it's random so you don't.
But for risk management it really gives you a much better sense of the fact
> you might have a current portfolio and do some sort of Monte Carlo simulation to see how risky that portfolio is.
That's exactly the use case.
> What sorts of strategies did your company generate and advise the hedge funds on?
So- a particular asset would be currently priced in a such a way to be contrary to what the yield curve would predict?
What it would come down to is it would be mis-priced relative to its peers on that yield curve, right? And backtesting would tell you, would confirm that time and again, that anomaly would revert back to something more normal, right? So you could make a lot of money by taking these reasonably sophisticated yield curve positions. Where you were long at two points of the curve, and short at two other points of the curve, and you get these funky butterfly of trades, right?
But they'd be sort of neutral to the level of the curve, and even neutral to the slope of the curve, but there's these third and fourth factors that would be in revert all the time.
> So, to take a short position in a fixed income, you would have to be a credit swap of some sort? Or how would you do that?
Many ways you can do that. Credit swap certainly does it.
> I just think of with the way you can trade only electronically, you never really think about the borrowing and the reselling.
Right. we used to worry about that stuff. Yeah repo rates, and all of that.
But yeah, that's been extracted away to some extent, right?
But yeah, that's how we used to do it.
> But, so you talked about looking at yield curves. Did you all do anything in equities?
Or was it just primarily the fixed income?
After 90s , And you could even do more sophisticated stuff like the stat arp stuff or eigenvalue PCA analysis on stock markets and do sort of payer trading strategies on these.
> Anyways, I wanted to ask to what do you attribute the evaporation of alpha and these kinds of approaches?
There's no question that the number of participants executing similar strategies is a factor.
> So if someone were, some young whipper wants to go out and start creating some kind of systematic strategy, is there any alpha left out there?
Or where should somebody like that start?
Oh, I'm sure there's all kinds of alpha to be had. It's just you have to, don't look in the same old places because everybody already has been there and there's only a bit.
I think there is all kinds of opportunity if you find new information sources.
You hear great stories about little creative ideas. And then you get a sense of how production is going for that company. Very simple little information advantage.
>Mechanical Turk like sort of things. The one thing I would say to sort of reinforce the kind of thing that Quandl can offer is we often find that the combination of different data sources adds value. So as an example, maybe over simplified. A combination of multiple technical indicators plus some fundamental data, putting all those together instead of just a simple, single factor model can work.
Yeah, I still have, there's three rules of thumb for building quantitative strategies these days, I think one should look for theoretically sound ideas. So don't bet on black box, neural net, kind of things, because they always fall down, right?
But if something makes theoretical sense in an economics or financial sense, it's a good chance you're onto something.
Number two is empirically tested, right? So build something that's theoretically sound and then show yourself empirically, with data, that this thing worked.
And then the final thing is be scared of complexity. If it doesn't stay simple its a warning sign your probably over fitting and getting yourself into a trap.
But if you devise a strategy that ticks three boxes, that it's sort of theoretically sound, empirically testable, and still simple, you're probably on the right back.
> One last question. So we've talked about things to do or how to attack it. What should people run from, what should they be scared of?
I think the single biggest trap, and I've fallen into this thing myself, right, is if you find yourself calibrating the same model on the same data usually, and you calibrate it.
It doesn't not work. And then you turn a few knobs and it doesn't quite work.
And then the eighth time you finally turn the knobs just right and it's working?
Odds are you've just curve fitted, right?
Don't fall into that trap of endlessly iterating on the same data.
Because you're in for unpleasant surprises once you get into, out of sample data.
0 개의 댓글:
댓글 쓰기