[UrmaBlume]

The need for speed
November 3rd 2008

Irene Aldridge, quantitative portfolio manager and managing partner at Able Alpha Trading in New York, looks at what defines a good high-frequency system

One out of every two money management job vacancies listed in October 2008 on the finance recruiting site eFinancialCareers.com was a search mandate for "high-frequency trading" professionals. A rush for a specific job category in the middle of the worst crisis since the Great Depression is unusual, particularly when most companies are enacting hiring freezes.

The reason for the hiring spree in the high-frequency field is simple: high-frequency trading is capable of generating money in all market conditions, whether a crisis or business as usual.

High-frequency trading is a new discipline among buy-side money managers. The main innovation is more frequent turnover of capital in response to changing market conditions. High-frequency environments are characterised by lower average gain per trade and a higher number of trades. While traditional money managers hold their trading positions for weeks or months at a time, high-frequency money managers typically execute multiple trades each day with few, if any, positions carried overnight. The absence of overnight positions is important for two reasons: 1) the continuing globalisation of capital markets extends most of the trading activity to 24-hour cycles, and with the current volatility, overnight positions are particularly risky; and 2) overnight positions taken out on margin have to be paid for at the interest rate referred to as an 'overnight carry rate', which is usually slightly above Libor. But with volatility in Libor and hyperinflation around the corner, overnight positions become increasingly expensive, and so unprofitable for many money managers.

While high-frequency trading enables the same strategy to be used across a wide range of financial securities, developing these trading strategies presents new challenges for money managers. The first fundamental challenge of high-frequency trading is the large volumes of intraday data. Most prudent money managers require at least two years of back-testing of the trading system to consider putting money behind it. Credible systems usually require four or more years of data to fully examine potential pitfalls, and dealing with this volume of numbers can be overbearing for most.

Another issue is that signals must be precise enough to work in fast-moving markets, where gains could quickly turn to losses if the signals are misaligned.

Speed of execution is critical to high-frequency trading. Traditional phone-in orders are not sustainable within the high-frequency framework. The only reliable way to achieve the required speed and precision is computer automation of order generation and execution. Programming high-frequency computer systems requires advanced skills in software development. Run-time mistakes can be costly and human supervision of trading in production remains essential to ensure the system is running within pre-specified risk boundaries. Discretion embedded in human supervision, however, should be limited to one decision only: whether or not the system is performing within pre-specified bounds and, if it is not, whether it is the right time to pull the plug.

What defines a good high-frequency system? As with any money management activity, the first metric to consider is the Sharpe ratio. For trading systems with no overnight positions, the Sharpe ratio equals the mean of returns divided by the standard deviation of returns. An annualised Sharpe ratio of 4, after all transaction costs, is becoming a de-facto benchmark for a solid, stable system in the industry. A system with a lower Sharpe might be profitable for short periods of time, but is statistically subject to blow-ups.

An annualised Sharpe ratio of 4 corresponds to a daily Sharpe ratio of 0.25. That is, daily standard deviation can be at most four times the average daily return. So, if the system or manager you are considering employing produces 0.1% per day on average, the maximum daily standard deviation of the returns should ideally fall under 0.4%. What this means in turn is that 68% of all daily returns should fall within one standard deviation from the mean, and 95% of all daily returns should fall within two standard deviations from the mean. In our example, 68% of all daily returns should fall within the -0.3% to 0.5% range, and 95% of all daily returns should fall between -0.7% and 0.9%.

The second consideration is sensitivity to latency. Many fast-moving markets are sensitive to timely execution. So a thorough understanding of what costs are involved when execution is delayed is a critical factor in understanding viability of a high-frequency trading system.

Other traditional metrics apply as well. A prospective investor in a high-frequency system should ask the system's manager questions about the maximum drawdown (a maximum peak-to-trough loss), betas (sensitivity to Standard & Poor's 500 and other macroeconomic indicators), value at risk (the loss potential at 95% probability level) and Sortino ratio (return over T-bills divided by average underperformance) among others. The manager's answers will not only indicate the stability of the system, but also reveal the manager's knowledge of, and attitude towards, risk management practices.

Overall, high-frequency trading is a difficult but profitable endeavour that can generate stable profits in various market conditions. Solid footing in both theory and practice of finance and computer science are the normal pre-requisites for successful implementation of high-frequency environments. And while past performance is never a guarantee of future returns, solid investment management metrics delivered on auditable returns net of transaction costs are likely to give investors a good indication of the high-frequency manager's abilities

[Frank]

This article is really about a concept written about in the 1990's about how to increase the reward-risk relationship.

There are 2 ways to make your trading better. Find strategies that increase your ‘edge’ (the mean 'expectancy' of a trading method over the longer-term)--- difficult to do as trading is a very competitive game and there will always be a cap on how high you can get your % win rate.

Second, and this is where the world is going -- find strategies that have solid edge but you can repeat them more (high frequency trading). The sharpe ratio increases with either of these 2 ways.

I wrote about this here:

http://www.traderslaboratory.com/for...gies-3603.html

The example I use is how to 'think' about strategies and taking the popular game of roulette as a useful example.

I learned this while reading a book called: "Active Portfolio Management" by Grinold & Kahn. This book is considered a bible by many money managers --- it is total overkill so I don't recommend the book unless you are a math major headed for quantitative finance. But the points made in it apply to all strategies, including short-term trading. If you compare two methods of trading, the one with lower edge might be a much more efficient strategy if it is high-frequency --- this is because the # of trades increases the statistical significance of the results and just like the house edge in roulette -- as the house, you would much rather do many spins at $25k each than just one roll at $1 million. In fact, you wouldn't even bother with the risk of $1 million, despite knowing you have an 'edge' in the game.

[BlowFish]

Hmmm intresting article but is it really new and innovative? Amongst "buy side money managers" perhaps. Guess its all new to Irene Aldridge.

An article I would have expected to read in the late eighties or early nineties (similar economic climate then too)

[AgeKay]

The Sharpe ratio that is mentioned in this article has one big flaw that many people do not consider: It penalizes big profits. So if you make 10% profit in a day where 1% is standard, the standard deviation will be huge and therefore the Sharpe ratio rather low. There are much better measures for consistency, but I don't want to get into that...

[sneo]

i agree, sharpe has a flaw due to them using standard deviation of returns (regardless of +/-ve). the industry uses sortino ratio.