Integrated Analytics' Droid Gives Traders Early Warning

Integrated Analytics, a Los Angeles-based expert systems software developer, is cooking up an intelligent alert system. "MarketDroid" is intended to emulate the behavior of a trader's assistant specially trained to watch market data.

The problem with existing alert systems is that they lack discrimination. Alarms go off too often or too late. Set the alert threshold too high and your entire screen is flashing; too low and you miss the boat. MarketDroid benefits from several earlier projects managed by David Leinweber, chief scientist at Integrated Analytics, for the likes of RAND and Inference Corporation.

PICON, a 'process intelligent control' tool he developed for the process industry was latched on to by numerous Wall Street quants looking to redesign it for financial applications. "But there are certain differences between securities trading and refineries that meant the tool had to change somewhat," says Leinweber.

A trader's alert demonstration developed at Inference Corp. succeeded in using pseudo natural language to express varying degrees of enthusiasm for long or short positions in specific stocks, "but the rules were hard-wired and the system was an incredible churner. It would have you in and out of the same stock five times before lunch," says Leinweber.

MarketDroid maintains market histories, reasons about what recent data mean in the context of that history, gives advice, and explains where the advice came from. MarketDroid uses:

A set of semi-autonomous agents that monitor market events and perform computations to evaluate derived indicators,

Rules to identify patterns of interest in direct and derived data, and,

Tools to allow customization.

The 'Droid can handle any number of securities. Alerts, which appear as anywhere from one to five up or down arrows, are displayed in a scrollable window of variable size. Higher level alarms automatically find their way to the top of the window.

Using Dempster-Shaeffer laws of combination of evidence and variable precision logic, the demonstration version of Marketdroid looks at the following data points:

last price vs. 90-day moving average,

volume over last 5 minutes vs. two-day average,

block volume vs. average block volume

position of last price relative to bid or ask

call volume relative to recent average

put volume relative to recent average

movement of competing stocks

options prices relative to theoretical values

put or call premium trend

Traders may input exceptions to their own rule base and special alarms are programmable. For better or for worse, the MarketDroid can gradually be trained to reflect any individual trader's interpretation of market events.