Each day, machines, not humans, are plowing through statistics and crunching numbers trying to devise trading algorithms to make money in stock markets, futures markets, commodities markets, and any other place that traders can turn a dime.
They’ve been successful.
But not totally successful. And not successful for long stretches.
The stories of trading strategies failing are just as common as stories of amazing machine-learning success in the market.
Why is that?
I’ll take a stab. I believe that most machine learning and artificial intelligence programs are essentially created short-sighted. To build an automated trading system, you “train” the program to understand data. This data can be technical or fundamental, or a range of other data sets. The program learns the relationships between the data and the market. When factor x goes up, the market reacts with a y, let’s say.
But, this data is not an object, per se, but is really a shadow of currents in a broader economy. So, the program ends up not be predictive at all. It is reacting to a reactive set of data.
Once market conditions change, like a gear turning a gear turning another gear, it will ripple throughout other sectors of the economy, affecting all the indicators and eventually shaking the set of data that has trained the automated trading program will begin to fail.
Markets are driven by many things — economic confidence, greed, fear, demographics, technological innovation, etc. — that do not show up in an indicator, no matter who supposedly predictive.
Most indicators are like signs that warn drivers of a dangerous curve after the motorist has already driven around the treacherous bend. For most investors, though, the indicators’ warnings come too late.
In order to be successful, automated trading programs will have to learn how to be adaptive, sensing changes to multiple facets that drive the market — deep facets, rather than responding to indicators that are actually nothing more than lagging warning signs.