Writer's Market

A couple of days ago, I encounter 2012 Writer's Market Deluxe Edition (Writer's Market Online) at a neighborhood bookstore. The book contained a catalog of freelancing opportunities in a wide assortment of magazines and journals. What I found interesting was that financial and technology freelancers did not receive an appreciably higher rate than other areas. There were certainly exceptions, but for the most part, there was no real discrepancy between rates in different fields. The main trading-oriented publication the book listed was www.traders.com (Technical Analysis, Inc., publisher of Stock and Commodities, Working Money, and Traders.com Advantage magazines), where freelancers apparently contribute a good deal of the articles.

According to the Editorial Guidelines, the top 5 hot topics are relative strength indicator (RSI), average directional movement rating (ADXR), Fibonacci ratios, MACD and moving averages, and volume trends. I am a bit surprised to see RSI so prominent. In contrast, stochastics and chart reading are much further down on the list.

A Paper on "High Frequency" Treasuries

There is a paper from London Business School dated from a while back (2002) that studies simple pairs strategies applied to the whole universe of US treasuries secondary market. The data used are intraday quotes and transaction information from GovPX (thus "high frequency"), which has since been acquired. The paper has an interesting table comparing equity and treasury markets (Table 1). One particular metric is the "social acceptance" of shorting the securities. The claim is that shorting equities has low acceptance whereas shorting treasuries is "less affected". The proposed strategies reduced the standard deviation of returns considerably though the level of returns were also dragged down somewhat. The best performing strategies generally kept risk controls very tight. When transaction and financing costs are accounted for, the returns start to trail the risk free return, though it outperforms equity and the treasury bond index in terms of Sharpe ratio and Gain-Loss ratio. Curiously, these strategies are slightly positively correlated with equities and slightly negatively correlated with the bond index.

Multiarmed Bandits

I tweeted a little while ago about a post from the Untyped people on applying a multiarmed bandit algorithm to website optimization (in contrast to A/B testing). It turns out that many have considered the multiarmed bandit problem for optimizing market maker profitability, Bandit Market Maker by Penna and Reid [PDF]. The basic idea is to dynamically optimize (i.e., reinforcement/online learning) the use of k-levers each operating under initially unknown and eventually partly known distributions. Presumably, the more one uses the levers, the more one learns about the reward, but there is an exploration-exploitation dilemma -- since we are operating online one has to choose whether to take the empirically revealed best choice thus far or to continue to explore for more rewarding options. The exploration-exploitation dilemma also matters to genetic algorithms and programming, but the multiarmed bandit problem explicitly optimizes for it. The exciting thing is that after decades of research, they have found optimal algorithms for the multiarmed bandit problem. The Untyped post cites one of them, Finite-time Analysis of the Multiarmed Bandit Problem by Auer et al [PDF].

Genetic Programming Performance

There is a whole lot of literature on the subject of genetic and evolutionary algorithms for trading strategies. The quality of the work seems to vary considerably. I have yet to discover the canonical and seminal lines of work in the field. Unlike in programming languages research and perhaps other scientific fields, the bibliographies do not all follow the same chain of citations. Or perhaps they do, but I have yet to discover it. More on this after the jump.

Genetic Programming and Trading Strategies

From an anecdotal sampling of Linkedin and web discussions, people have applied various machine learning techniques to the creation of trading strategies. The most popular techniques for trading strategies mirror that of the general applied machine learning community, meaning SVM and neural networks are frequently mentioned. One relatively lesser known technique is genetic programming (GP), a variant of genetic algorithms for tree data structures, normally abstract syntax trees. Curiously, the academic community has studied genetic programming quite a lot. Indeed, there are many papers on genetic programming for trade strategies in GECCO (a conference on genetic and evolutionary computation) and other conferences/workshops. Some people from State Street have even contributed their study. The Quant Stack Exchange has a discussion on genetic algorithms where the problem of data snooping is frequently mentioned as a shortcoming, a problem addressed by regularization in conventional learning techniques. I informally survey some of the studies of the effectiveness of GP in a follow-up post.

My cursory interest in this area is the potential impact of programming language design and especially static type systems on the performance of genetic programming for trading strategies. The nice thing about genetic programming is that the results produced can be theoretically unrestricted unlike basic machine learning techniques that make assumptions such as linearity about the underlying model. Practically, however, the results produced by genetic programming are usually small and possibly uninteresting. This is where modern programming languages theory comes in. In the past decade or so, a number of researchers have studied the problem of the generation of random test cases (of which QuickCheck is a prime example). One of the main engineering challenges there is to ensure that the test cases generated do not all fall into a trivial category. The test cases here can range from instances of simple primitive types (e.g., ints) to nontrivial data structures (e.g., syntax trees or data packet queues). Usually it falls to the developer to write functions that generate a good distribution of test cases to ensure, for example, that the probability of getting a list of length greater than 3 is not negligible. In practice, QuickCheck supplies an assortment of combinators that help produce a good distribution. The developer's domain knowledge is still necessary, but the test generation problem reduces down to a choice of combinators.