Friday, October 10, 2014

Cython is awesome




After some initial frustration with learning Cython, I found the Scipy 2013 Cython tutorial by Kurt Smith of Enthought, which was a hugely positive surprise. IMO it's by far the best intorduction to Cython (the official Cython tutorials can be a bit tedious...).

So I decided to put my newly learned skills to use in optimizing some stochastic differential equation simulations. To make a long story short, here's the original Python+Numpy version:


...and here's the Cython optimized version:
Previously I had to wait about 30s for 1000000 iterations, and now it takes only about .1 seconds... definitely worth the effort ;) (BTW sorry about the lack of documentation... anyway, the algo has an adaptive time step, since I'm simulating some pretty wildly behaving SDEs).

Wednesday, August 6, 2014

Holographic principle

So just a while ago I received an email from World Scientific about my article "Easy Holography", which was published in their journal IJMPD. According to the email, my paper "...is a great contribution in understanding the relevant field of study, which is worth to be shared." Oo well gee, thanks for the recognition World Scientific (I'm sure only select few receive such encouragement)! Oh well, I guess I'll just do as I'm told and share the article!

Anyway, the idea is roughly as follows (it's been a while since I worked on this stuff, so my apologies for some possible inaccuracies): let's pretend we live in a two dimensional universe and that this universe is a two dimensional quantum gravity theory, which is "asymptotically $AdS_2$", which is very similar to a hyperbolic plane $\mathbb H_2$ but of minkowskian signature (google it or stop reading). Or just think of $\mathbb H_2$. The "asymptotic symmetries" of $\mathbb H_2$ are such diffeomorphisms which preserve the metric near the boundary. For $\mathbb H_2$ these symmetries are isomorphic to circle diffeomorphisms $Diff(S^1)$, which is an infinite dimensional group. The only thing we need to assume about the quantum gravity theory is that it is integrable w.r.t to the asymptotic symmetries, which in this context just means that every physical state can be mapped to another physical state by the asymptotic symmetries extended to all of the space (a la Conformal Field Theory).

The toy model I used in the "Easy Holography" paper was basically just a pixelated image (of a bunny, since it was submitted on Easter). A rough estimate of the entropy of that "model" is then a logarithm of the number of all distinct images that are mapped from an original image by the asymptotic symmetry group, which is isomorphic to $Diff(S^1)$.


So first I have a bunny. Now "coarse grain" it on an $N \times N$ grid, so that a grid is black if most of the (Euclidean) surface area under it is black and white otherwise. Then apply a map from the asymptotic symmetries on the original image and coarse grain that. If the coarse grained image is different from the original coarse grained image, count that as a distinct state/configuration. Repeat this for all as. symmetries (there's a noncountably infinite number of them) and write down the number of all distinct images, or configurations, as a function of the number of  $N$.

If the maps were just the usual diffeomorphisms, it's easy to see (see the image above or the paper) that the number of distinct configurations is just $2^{N^2}$ and thus the entropy is $S = N^2 \log(2)$. So the entopy is just proportional to the volume of the system, $N^2$. So that's the old physics. But if instead we use the asymptotic symmetry maps, it turns out that the entropy thus computed is now proportional to $N$, or the "area" of the system! This happens also for $AdS_3$, whose as. symmetries are the conformal group, isomorphic (if I recall correctly) to $Diff(S^1) \times Diff(S^1)$. In the $AdS_3$ the entropy would be proportional to $N^3$ (volume) for ordinary diffeomorphisms and to $N^2$ (area) for the asymptotic symmetries.

So the conclusion would seem to be that the asymtptotic symmetries and the asymptotic $AdS$ structure of the universe is somehow related to the Holographic Principle, which states that the entropy is indeed proportional to the surface are of the system, instead of the volume... but of course the universe is not three dimensional! So what happens in higher dimensions? The asymptotic symmetry group of $AdS_4$ is disappointingly not infinite dimensional, but instead the six dimensional $SO(2,2)$. That would actually just lead to a logarithmic entropy, so no luck there...

But what actually is an $AdS$ space and why would they be interesting? Let's look at (one) definition of an $AdS$ space. We can define $AdS_d$ as the homogeneous space $SO(2, d-1) / SO(2, d-2)$. For $d=2$, we have $AdS_2 \simeq SO(2, 1) / SO(2)$ and for $d=3$, $AdS_3 = SO(2, 2) / SO(2,1)$. For low dimensional Lie groups there exist so called "accidental isomorphies", which we can use to write $SO(2,1) \simeq SL(2)$ and $SO(2,2) \simeq SL(2) \times SL(2)$, where $SL(2)$ is the special linear group. So we can write the $AdS$ spaces in an equivalent way as $AdS_2 \simeq SL(2)/SO(2)$ and $AdS_3 \simeq SL(2)$ (Lie groups are manifolds too). For $d>2$ such accidental isomorphisms do not exist. More specifically, the higher dimensional $AdS$ spaces are not related at all to the special linear groups $SL(n)$.

A short recap: $AdS_2$ and $AdS_3$ are intimately related to the Lie group $SL(2)$, but higher dimensional $AdS$ spaces are not. $AdS_2$ and $AdS_3$ admit infinite dimensional as. symmetries, but higher dimensional $AdS$ spaces do not. Clearly the Lie group $SL(2)$ is the key here! Some natural questions are then

1) Do the higher dimensional Lie groups $SL(n)$ (or related homogeneous spaces) also admit infinite dimensional asymptotic symmetries?
2) If they do, will they also incorporate a "natural holography" as with $AdS_2$ and $AdS_3$?
3) Is it then possible come up with a higher dimensional quantum gravity theory, which is similarly exactly solvable as $AdS_3$ quantum gravity (well, it should be exactly solvable I guess...)?

The answer to part 1) above is "yes". I actually showed that for a homogeneous (symmetric) space $SL(3)/SO(3)$ in an earlier paper called "Infinite symmetry on the boundary of $SL(3)/SO(3)$". That paper is not very elegant, and I was working on an extended and more sophisticated version, but I didn't have time to finish writing it... the final equations were actually solved by Robert Bryant at mathoverflow. I couldn't get any research funding for that though and now I'm doing something completely different! Anyway, hopefully someone will some day find these results useful.






Wednesday, December 18, 2013

Statistical arbitrage with Python

UPDATE: The current Github version of the backtest is a bit broken: there was a silly bug that caused the algo to "see" 1 min into the future. Also, the overnight effects/ jumps in the morning kinda ruin the intraday trades, so I'm currently rewriting the code for 1 sec bar bid/ask data to be used intraday only... the current model is also a bit too naive (especially for higher frequencies) and will need several other improvements to be useful in practice. I will probably not share it publicly in the future though, but if you want to talk about it, feel free to drop me an email etc.

Finally managed to complete an early version of my PyArb statistical arbitrage project... I published it in GitHub as a Python module here, although the best way to view it right away would be to check out the IPython Notebook here at nbviewer.ipython.org. It is a model dependent equity statistical arbitrage backtest module for Python. Roughly speaking, the input is a universe of N stock prices over a selected time period, and the output is a mean reverting portfolio which can be used for trading. The idea is to model "interacting" (correlated, anticorrelated or cointegrated) stock prices as a system of stochastic differential equations, roughly as

$$ dX_t^i = A^i_j X_t^j dt + X_t^i dW_t^i,$$

where $X_t$ are the prices and $dW_t$ are white noises.

The stochastic part doesn't yet play any important role, but that will soon change...

This is just a backtest for a strategy, so there's no saying it will actually work in a live situation (but I'm planning to try paper trading next). Specifically, there's no slippage and impact modelling, short sell contract and borrow costs etc. I just assumed a flat rate \$.005 per share cost from Interactive Brokers' website as a sort of ballpark figure. It gives a roughly 12% annualized returns with a Sharpe ratio of about 5 and a maximum drawdown of 0.6%. Maybe that sounds a bit too good to be true? Well maybe I made a mistake, go ahead and check the code! :) (I need to check it again myself anyway or give it a go in e.g. Quantopian).

Here's a plot of the cumulative returns for a period of about 300 days. The "mode=0" is the best portfolio and corresponds to the lowest eigenvalue of the evolution matrix $A$ in the equation above.




Saturday, September 14, 2013

Funny statistics in stock market data

Funny statistics in stock market data

Market data structure functions

I recently got some minute level stock market data from The Bonnot Gang for some data analytic (and stat arb design) purposes, when I noticed some funny behavior in the data structure function. Now the concept of a structure function may not be very widely known with quants/ data analysts/ economists, so here's a definition:

Suppose there's a time series Xt. The structure function of Xt is defined as

Sn(τ)E(|Xt+τXt|n)

where for a given sample of data you just replace the ensemble expectation E() by the sample mean, 1Nt=0N().

These types of structure functions have been studied for some time now in finance in the context of similarities between financial markets and hydrodynamic turbulence. I think it all started in 1996 with the paper Turbulent cascades in foreign exchange markets by Ghashghaie et al. They computed the structure functions for some FX market data, and found a scaling relation Sn(τ)ξn, where ξn is a concave function of n, implying multiscaling in FX markets, similarly to hydrodynamic turbulence (BTW their conclusions about the result were a bit out there, but I guess the data analysis is still good).

So I did some of my own data analysis with the Bonnot Gang data (I hope it's not bad data!). Here's a few plots of the structure functions, first for n=1:

$n=1$

Then n=3:

$n=3$

This is close to linear, i.e. ξ31, as in turbulence. Then n=10:

$n=10$

Clearly you can't fit a power law in all of this, but there seems to be clear power law regimes divided by about 6, 18, 60 and 180 minutes! I don't know the reason for this, but if I had to guess, I'd say it's because of traders/ algorithms operating w.r.t different data timeframes... or maybe it's because of the finite tick size...

Anyway, I don't have time to get to the bottom of this, but maybe someone else will... so if you see this stuff on a paper someday, you saw it first here!! ;)

Written with StackEdit. Try it out, it's awesome!! You can do MathJax and sync everything in Google Drive or Dropbox and publish directli in Blogger, Wordpress, Tumblr etc.!

Tuesday, April 30, 2013

Goodbye to Academia

Damn I'm tired of applying for puny research grants (and not getting one most of the time)... Time to move on and implement plan B that I've been preparing for some time: I've decided to squeeze myself into quantitative finance! It's actually really interesting. I was greatly inspider by Emanuel Derman's "My life as a Quant", where he describes pretty similar circumstances to my own with his postdoc aspirations. I recommend the book to anyone considering leaving Academia.

From a mathematical point of view, I was a bit awe-struck by Pierre Henry-Labordere's "Analysis, Geometry and Modeling in Finance". Apparently there's a close connection between the geometry of black holes (well, at least the hyperbolic types) and stochastic volatility models! It would be extremely cool to be able to actually research and design models like that which would actually work and be useful! Anyway, I doubt real quant life is that fancy, but I'll be happy if the job involves PhD level math of some kind.

I'd love to hear other people's stories about switching from academia to industry or finance. Anyone here with similar experiences?

Friday, November 2, 2012

Workshop: LIE THEORY AND ITS APPLICATIONS IN PHYSICS

I received an invitation to the conference "LIE THEORY AND ITS APPLICATIONS IN PHYSICS", which will be held in Varna, Bulgaria next june. That would be my first hep-th/gravity related conference... exciting! Some interesting names in the advisory board and previous years' visitors. Now if only I could get some funding, I could actually go (a hint for possible reviewers reading this).

Sunday, September 30, 2012

An idea for gauge/gravity dualities

Change of plans. Converting TeX to HTML and the post editing involved is a drag, so I'm not going to finish the series about asymptotic geometry etc. Instead I wrote a new paper on the subject! The preprint is posted on arXiv here, but I haven't submitted it yet(I will probably revise a bit). Here's a short summary.

In the paper I try to generalize the successful results of gauge gravity duality for $AdS_3$ gravity to higher dimensions (or at least motivate the attempt). This was the big idea by Brown and Henneaux (link to paper at inSPIRE.net) which was later refined in here (and in other people's works, but I don't want to cite a lot of sources). The idea is that a three dimensional gravity theory that approaches an anti-de Sitter space at a spherical boundary at infinity is equivalent (or dual) to a two dimensional conformal field theory on the boundary, which turns out to be a 2D Liouville theory. Brown and Henneaux discovered this by studying the asymptotic symmetries of $AdS_3$ gravity, which they discovered to actually tend to conformal transformations at the boundary. The asymptotic symmetries are here diffeomorphisms which preserve the asymptotically $AdS_3$ conditions. Moreover, they discovered that the algebra of canonical generators of these asymptotic symmetries is the Virasoro algebra with a classical, nonvanishing central charge! 

What's perhaps even more interesting is that this kind of approach does not work in higher dimensions. The reason is that the boundary of $AdS_{d+1}$ is a $d$ dimensional flat space (Minkowskian or Euclidean, depending on if you're talking about real $AdS$ spaces or the Euclidean versions, i.e. hyperbolic spaces) and because the conformal transformations in dimensions $d>2$ forms a finite dimensional group. The infinite dimensionality of the asymptotic symmetries will be the key!

As I explain in the paper, the key to understanding how to generalize the extremely successful 3D case to higher dimensions is the observation that (loosely speaking) the 3D $AdS$ space is in fact the Lie group $SL(2)$. Indeed, my hypothesis was that instead of studying higher dimensional $AdS$ spaces, one should study the geometry of the Lie groups $SL(N)$ (and their homogeneous spaces). In the first article of the subject, (published in JMP),I discovered that the asymptotic symmetries on the boundary of the five dimensional homogeneous space $SL(3)/SO(3)$ form an infinite dimensional Lie algebra, which implies the generalization of the duality (the paper was a bit of a "tour de force" and not very elegant...). This is completely different to the higher dimensional $AdS$ spaces, where the asymptotic symmetries form a finite dimensional Lie algebra. It will then be possible to extend the $SL(3)/SO(3)$ asymptotic symmetries to near boundary symmetries to obtain a similar dual theory as the Liouville theory in the 3D AdS gravity case.

Anyway, I hope the stuff makes sense... I wrote this piece quite quickly. The intro in the paper is probably better!