TRADETECH WEST: Next, Algos That Think for Themselves
September 11, 2012
SAN FRANCISCO -- As mathematical formulas go, the algorithms used by brokers to execute trading strategies so far are fairly straightforward.
They define a target quantity of a stock to sell or buy and try to spread the impact of the trades that result, over time.
That could be about to change. The Algorithmic Trading Management unit of Cowen Group, for one instance, described the processes and early results of algorithms it is developing that learn as they go, from what they experience in trading in a given market.
This will lead to, in the estimate of chief executive Doug Rivelli, “algos that can really think for themselves and think on the fly.’’
The company’s approach used a series of tests of “micro” strategies, in five-minute trading windows, to determine how successful they could be in achieving better prices in an overall “macro” trading strategy.
ATM used, according to its director of quantitative development, Jack Gold, these mathematical approaches to determine whether its approach could accurately predict the upward or downward movement of a stock.
• LOGISTIC REGRESSION. This is a mathematical model that predicts the movement of one variable based on other independent variables. By relying on, in its test, six months of historical data, the approach would measure the impact on the dependent variable – a stock’s price – from a variety of continuously changing variables, from price movement to order book data.
• SUPPORT VECTOR MACHINE. This is an approach that models key factors as points in space, tries to maximize the separation of the observations and then predict the result that the next piece of data will produce, depending on which side of the gap it falls on.
• RANDOM FOREST. An ensemble method of analysis that uses decision trees to produce a series of predictions about what will happen next. The approach then averages the predictions, to come up with a final prediction.
All of which happens in tiny slices of time, fast enough to compete in a trading environment where trades are won by how fast new orders get to market. The winning time is now measured in millionths of seconds, typically.
The models are trained on six months of data and constantly recalibrated by the five-minute windows they are run in.
At this point, they don’t try to predict how much a stock will move up or down, just that it will move up or down.
In its early tests, Gold said the Random Forest approach produced the best results. That technique was correct 58 percent of the time, when it got a strong signal to act. And it produced a 1.75 basis point improvement in the price obtained.
This, he said, is a demonstration of the "power of machine learning techniques for prediction.’’
To date, brokers have not likely used machine learning in any significant manner, at this point, Gold said. But some mathematically driven trading firms, such as statistical arbitrageurs, do make use of the practice.