Sunday, March 13, 2011

Second Iteration

Darlene had been a great help. She someone who not only had been at Carnegie-Mellon’s language project, but also could conceptually reverse-engineer Watson’s algorithms.

Garland’s decision to hire crack programmer Iain Quan was almost instantaneous. A refugee from Carnegie-Mellon, Quan had already mapped out a plan built on daily gold reports on the Internet. Tracking gold with the finest-grained inter-day data Enlightened could afford, the news reports throughout the day would be integrated with the metal’s movements. Word-matching would be used to process details on the varied explanations for why gold rose and fell. It required a large news database, but that was okay. Curious, Iain had looked at the goldbug world and became more interested. Some of his school chums, nice guys, had talked like those people. After stumbling upon the “Original Gold Bug,” Iain had christened the new project DiNES: The Distinguishing Natural English System. Like any computer professional, he had thought beyond the specific purpose for which he had been contracted.

Despite copious use of opens-source modules, a community chest to which he himself had contributed, it took four months for him to get DiNES 1.0 off the ground. One more month, and he had 1.34 and something to show the top boss.

Said Harold Garland visited the temporary workspace at the foot of the conference table. The table was largely for show, and he was pleased to show a hard at work top-of-the-line programmer to visiting clients. Had Iain been on salary, he would be pulling down north of $50,000 per month. He was that good, and that needed. As it was, he was being paid by the job with advances every month.

By 1.34, it was a job well done. Testing it took yet another month, and not because Iain Quan wanted to bleed anything out. His $500,000 bill was fixed no matter how long it took. The month was needed for a thorough real-time test, including a Turing test.

Sure enough, the journalists on the gold beat treated DiNES as a soon-to-be-distinguished cubbie sans sources. Not knowing that they were corresponding with a computer program, because Iain himself handled the E-mailing on DiNES’ end, they not only were beat by “Sandy Grey” but they also started copying “him”. Iain wondered if a “Sandy Grey” module should be added to the DiNES project. He decided against it, reluctantly, because it would take too much time. Enlightened Capital wasn’t a journalism outfit; it was an investment firm. DiNES 2.0’s job would be to find algorithms that could tease out profitable trading strategies for gold. Although a fruitful path for the concept, a full Turing-test module didn’t work for the client.

Iain was surprised to see two kids with his employer. The boy was clearly Mr. Garland’s son; by inference, the older girl was his daughter. She had dark eyes while the boy’s were blue; her mouth was thinner than her younger sibling’s. As was her body.

Running through the demonstration, Iain carefully deployed the E-mail correspondence. After showing some real-time samples, he asked DiNES to print out an on-the-spot report. “Show it to a gold trader,” he offered.

The three went away, with the boy sporting an excited grin. It only took a couple of minutes for the report to be scanned, E-mailed, read, and responded to by the head of gold trading at a friendly full-service firm. He wondered to Harry what an executive partner would be doing with a journalist. “A good one – caught something I didn’t think of – but why the reporter?” Garland responded that Enlightened was getting into gold, but in a non-controversial way. Journalism seemed the best place to start.

Iain saw a pleased middle-aged man with a smiling daughter and a visibly excited son. “It works. It really does,” said Garland the genuinely satisfied client.

His son was far more voluble. Walking over beside Iain, his chatterboxing showed a solid conceptual knowledge of natural language processing. Little Hal was so assertive, his big sister took offence. “Don’t mind her, she’s stupid,” he said placidly to his new big buddy.

“I’m warning you,” she snapped back while whipping out her cell phone, “you’re going to get cybersmacked!”

“Like I care.” Brandishing his own smart phone, he quickly texted a message with “mr ltr” at the end.

Iain was fitting in. He found more and more people stopping by to see if DiNES 2.0 was up and running. Not wanting to bother the programmer while programming, they stepped in and out until the testing was being done. When he was finished the to-be-refined guts of the trading algorithm, he began attracting a crowd.

DiNES 2.21 was a bit of a disappointment. Statistical analysis showed no consistently profitable trading strategies, despite the enticingly patterned chart. Backtesting, using virtual real time, showed more promise. Since the program didn’t know whether it was 2012 or 1912, there was no way it could "backcheat" unless programmed to. It ran though thirty years of real time in minutes.

What it ended up with was a model that was an elaborated hybrid of multiple-regression decision analysis and multivariable arbitrage pricing theory. Some of the elaborations were qualitative instead of quantitative, although the final values and weightings all had to be real numbers. Thankfully, one of the stable components of the model was the level of real interest rates. The delta of the inflation rate was also there, but had less significance. Money supply changes flickered in and out. DiNES had come up with a crisis model that looked for headline news with words associated with disasters, and had sifted through to find ones that would move gold. There was a grab bag of others, including (interestingly enough) shorter-term statistical tests as compared with longer-term ones. In the virtual real time test, DiNES 2.21 became pretty good at getting into gold when it was consolidating. It had beaten buy-and-hold, and had largely avoided the 1980-2001 bear market. After a basic search, Iain was amused to see DiNES treat the “Brown bottom” gold sale cynically.

At first, he had unveiled to visitors a full-tooth smile that busy people use when they want to be polite to guests. As the testing continued, he became more relaxed. The project was in the bag once the real real-time testing was through.

No comments:

Post a Comment