On Jim Simons, String Theory, and Quantitative Hedge Funds

Renaissance Technologies founder and mathematics professor Jim Simons is an enigma in quantitative hedge funds.

 

Simons rarely gives interviews. One of the best is an Institutional Investor interview he gave in 2000 (PDF). One insight is that Renaissance makes trades in specific time periods using pattern recognition to model volatility.

 

Simons has done important work in differential geometry and the theoretical physics subdiscipline of string theory. I recently looked at some academic journal articles by Lars Brink (Sweden’s Chalmers University of Technology) and Leonard Susskind (Stanford University) to try and understand how Simons views financial markets.

 

String theory proposes one-dimensional objects called strings as particle-like objects that have quantum states. String theory and cosmology has progressed over the past 35 years to describe this phenomena but still lacks some key insights.

 

How might Simons use string theory to understand financial markets? Two possibilities:

 

(1) The mathematical language of couplings, phase transitions, perturbations, rotational states, and supersymmetries provides a scientific way to describe financial market  data and price time-series. It does so in a different way to fundamental analysis, technical analysis, and behavioural finance: Simons uses string theory to understand the structure of information in financial markets. (Ed Thorp pursued a similar insight with Claude Shannon using probability theory.) String theory-oriented trading may be falsifiable in Karl Popper’s philosophy of science.

 

(2) String theory provides a topological model that can be applied to money flows between mutual funds, hedge funds, and bank trading desks over short periods of time. This might enable Simons’ traders to forecast the likely catalysts for changes in stock prices in the short-term and to trade accordingly. This might involve using string theory to forecast how price trajectories might change if portfolio managers at other funds alter their portfolio weights for a stock. In doing so, Simons is trading in a similar way to SAC’s Steve Cohen (who uses game theory) and D.E. Shaw’s David Shaw but uses different methods of pattern recognition to do so.

 

I have made a list of popular science books and Springer academic monographs to keep an eye on string theory. Simons’ success also illustrates how insights from one knowledge domain (string theory, astrophysics, computational linguistics, and voice recognition) can be transferred to another domain (financial markets trading).

23rd June 2013: Google’s DistBelief & Market Mental Models

For the past several months I’ve been learning about machine learning and algorithms to build a quantitative trading capability.

 

Jeff Dean and his team at Google Research have developed DistBelief: a deep network that uses unsupervised learning, thousands of computers, and billions of parameters to train new models:

 

Recent work in unsupervised feature learning and deep learning has shown that being able to train large models can dramatically improve performance. In this paper, we consider the problem of training a deep network with billions of parameters using tens of thousands of CPU cores. We have developed a software framework called DistBelief that can utilize computing clusters with thousands of machines to train large models. Within this framework, we have developed two algorithms for large-scale distributed training: (i) Downpour SGD, an asynchronous stochastic gradient descent procedure supporting a large number of model replicas, and (ii) Sandblaster, a framework that supports a variety of distributed batch optimization procedures, including a distributed implementation of L-BFGS. Downpour SGD and Sandblaster L-BFGS both increase the scale and speed of deep network training. We have successfully used our system to train a deep network 30x larger than previously reported in the literature, and achieves state-of-the-art performance on ImageNet, a visual object recognition task with 16 million images and 21k categories. We show that these same techniques dramatically accelerate the training of a more modestly- sized deep network for a commercial speech recognition service. Although we focus on and report performance of these methods as applied to training large neural networks, the underlying algorithms are applicable to any gradient-based machine learning algorithm.

 

DistBelief has several interesting implications. It might underpin a new wave of consumer-oriented apps which use a machine learning architecture. It will likely discover new combinatorics and permutations of model parameters. Applied to historical and real-time data from financial markets it might discover new model parameters, usable for quantitative trading. This would potentially change our mental models of financial markets in an evolutionary way just as high-frequency econometrics disrupted indicator-based technical analysis models over a decade ago.

18th June 2013: Algorithmic Trading Goes Retail

Fortune Magazine reports that EquaMetrics is now selling a cloud-based app that creates Technical Analysis-based algorithmic trading strategies for retail trading subscribers:

 

EquaMetrics’ app is simply designed and since its software firepower comes from the cloud, it doesn’t require anything more than the typical PC. You can drag and drop colored tiles to assemble your own algorithm. Day traders can choose between 30 variables to build their formulas. The options are built on so-called technical indicators, metrics that reflect trading patterns as opposed to stock fundamentals such as the price-earnings ratio. After you’re done, you run the program to buy and sell stocks and currencies.

 

The web application is relatively inexpensive: it costs $99 a month or $250 a month, depending on how many algorithms you want to run. That’s a steal compared to the alternative of hiring a quantitative programmer for $200,000 a year. EquaMetrics gives you the stuff a programmer could produce. Then it’s up to you to assemble your own strategy.

 

I have been expecting apps like this for several months, and have been monitoring other initiatives like the Quantopian community. The popular literature on algorithmic trading strategies evolved from Technical Analysis mechanical systems (Tushar S. Chande’s Beyond Technical Analysis) to back-testing (Robert Pardo’s The Evaluation and Optimization of Trading Strategies) and then to algo trading using Matlab software (Ernie Chan’s Quantitative Trading and his new Algorithmic Trading; and Barry Johnson’s Algorithmic Trading & DMA). This period spans the post-dotcom collapse; the 2003-08 speculative bubble in real estate and asset-backed securitisation; and institutional experimentation with high-frequency trading platforms, and transaction and execution costs.

 

EquaMetrics’ strategy reflects this decade-long evolution:

  • Its initial offering is Technical Analysis strategies: at a time when: (a) high-frequency trading has ‘broken’ many trend-following and momentum indicators; and (b) hedge funds and proprietary trading desks use predatory trading to clean out TA-oriented retail traders.
  • The model is subscription-based software as a service — which could eventually disrupt or change the economics of agile software programming if this offering scales up in a significant way. Will the $US99-250 per month price point remain? Or will another platform develop a lower-priced offering and trigger a ‘race to the bottom’ competitive dynamic?
  • It opens the way for the licensing of specific TA indicators and proprietary methods as ancillary revenue streams, and as a way to build a market around the core product offering (which NinjaTrader, MetaStation, and ESignal have all done with their respective platforms).
  • The quality and scope of the back-tested data is important: quantitative hedge funds like Jim Simons’ Renaissance and David Shaw’s D.E. Shaw & Co each clean their own data.
  • EquaMetrics’ move into fundamental indicators reflects some recently published work on the quantitative analysis of these strategies (notably, Richard Tortoriello’s Quantitative Strategies for Achieving Alpha, and Wesley Gray and Tobias Carlisle’s Quantitative Value).
  • EquaMetrics’ choice of FXCM and Interactive Brokers as prime brokers to process client trades is significant: brokerage transaction and execution costs can mean a potential, new trading strategy is actually unprofitable to execute, or that its profit-taking ability declines over time, especially in correlated and ‘crowded trade’ markets.
  • The focus on TA and fundamental indicators does not address some of the quantitative, statistical or machine learning strategies that quantitative hedge funds use to develop algorithms; how correlation testing of model variables might occur; and what might happen to retail investors once several different competing firms have back-tested and issued dueling algorithms (a factor in high-frequency markets where scalping and order front-running occurs).

 

Still, the EquaMetrics offering has me interested: I’ve been waiting for algorithmic trading to ‘value migrate’ (Adrian Slywotzky) to retail traders, for awhile. It’s a first step towards post-human trading (Charles Stross’s novel Accelerando).