A decade ago, at a NY conference, an analyst put up slides showing his model of the short-term oil price (variables like inventories, production and demand trends, and so forth). I turned to the colleague next to me and said, “I just want to ask him, ‘How old are you?’” I worked on a computer model of the world oil market from 1977, when the model was run from a remote terminal and the output had to be picked up on the other side of campus. (Yes, by dinosaurs.) Although I haven’t done formal modeling in recent years, my experiences might provide some insight into the current fashion for using computer models in investing (among other things).
About two centuries ago, Baron von Maelzel toured the U.S. with an amazing clockwork automaton (invented by Baron Kempelen), a chess-playing “Turk” in the form of a mannequin at a desk with a chess board. The mannequin was dressed up as a Turk, given perceptions at the time of their perceived superior wisdom. The automaton could not only play chess very well, but solve problems presented to it that experts found difficult. Viewers were amazed, given the complexity of chess, and the level of play was not matched by modern computers for nearly two centuries. None of the Turk’s observers could initially explain the mechanism by which such feats were performed.
This is reminiscent of the 1970s when Uri Geller made claims to have paranormal abilities, which physicists from SRI found they could not explain. Because he wasn’t committing acts of physics, but sleight of hand, as demonstrated by the Amazing Randi who was not a scientist but rather an expert in the latter craft. (Similarly, peak oil advocates are often amazed at techniques done by scientists that are actually statistical in nature—and done wrong.)
Edgar Allan Poe considered the case and proved to be the Amazing Randi of his day. The chess-playing Turk was the result of “the wonderful mechanical genius of Baron Kempelen [that] could invent the necessary means for shutting a door or slipping aside a panel with a human agent too at his service…” in Poe’s words. He noted that the Baron would open one panel on the desk, show no one behind it, close it and open the other, again revealing no human agent; but this is just a standard magician’s trick, where the subject simply moves from one side to the other. Indeed, others claimed to have seen a chess player exit the desk after the audience had left.
Computer models often fall into this category. No matter how scientific and objective they appear, there is always a human agent behind them. In oil market modeling in the 1970s, this took the form of the price mechanism. NYU Professor Dermot Gately had suggested that prices moved according to capacity utilization in OPEC, as in the following figure (later used by the Energy Information Administration, among many others). If utilization was above 80%, prices would rise sharply, below 80% they would taper off.
This made sense, given that many industries use a similar conceptual model to predict inflation: high utilization in the steel industry results in higher steel prices, etc. And the model certainly seems to fit the existing data.
At least until 1986. After 1985, the data points no longer fit the curve, and the EIA stopped publishing the figure after 1987; for the last two years the model was well off. Subsequently, EIA ceased to publish the figure, although they used the formula for some time to come.
What had become obscured by the supposed success of the formula was that it was intended for use of short-term price changes. High steel capacity utilization would mean higher steel prices, but lead to investment and more capacity, so that prices would stabilize and even drop.
But oil models couldn’t capture this, because much of the capacity was in OPEC and it was assumed that OPEC would not necessarily invest in response to higher prices. Instead, the programmer had to choose numbers for OPEC’s future capacity and input them into the machine, meaning the programmer had control over the price forecast by simply modifying the capacity numbers. Despite the ‘scientific’ appearance of the computer model, there really was a man in the machine making the moves.
People have long sought to reduce the influence of fallible humans, whether replacing workers with machines or putting control of our nuclear weapons in the hands of Colossus, a giant computer that would avoid an accidental nuclear war. (1970 movie Forbin, the Colossus Project, fourteen years before Terminator’s Skynet). Ignoring that there is a always human element, even if only in the design.
Without any expertise in the field of artificial intelligence, it nonetheless seems to me that AI trading programs might learn, but won’t they learn at they are taught to do? Will this not simply be an extension of algorithms used by others in the financial world, at whose core as simply comparison of current with historical data and trends?
And this, after all, is what led to the financial meltdown described so aptly in When Genius Failed, the story of Long Term Capital Management and the way it nearly crashed the world economy. Recognizing patterns of behavior preceding an OPEC meeting, such as the way prices move in response to comments by member country ministers, can be useful, but will novel cases such as the SARS epidemic or the 2008 financial crisis catch the programs flat-footed, possibly triggering massive losses?
The answer, as it often does, comes down to gearing. LTCM’s model failed, but the problem was the huge amount of money they had at risk, far outreaching their capital. For a few small traders to use AI programs, or an investment bank to risk a fraction of its commodities’ funds would not be a concern. But if such programs become widespread, and they all programs are drawing the same conclusions from historical data, could there be a huge amount of money making the same bet?
For individuals, of course, the answer is diversify, one of the first investing lessons. I wonder how many AI programs will practice the same.