Symbols Without Context

Humans have invaluable real-world knowledge because we have accumulated a lifetime of experiences that help us recognize, understand, and anticipate. Computers do not have real-world experiences to guide them, so they must rely on statistical patterns in their digital data base—which may be helpful, but is certainly fallible. We use emotions as well as logic to construct concepts that help us understand what we see and hear. When we see a dog, we may visualize other dogs, think about the similarities and differences between dogs and cats, or expect the dog to chase after a cat we see nearby. We may remember a childhood pet or recall past encounters with dogs. Remembering that dogs are friendly and loyal, we might smile and want to pet the dog or throw a stick for the dog to fetch. Remembering once being scared by an aggressive dog, we might pull back to a safe distance. A computer does none of this. For a computer, there is no meaningful difference between dog, tiger, and XyB3c, other than the fact that they use different symbols. A computer can count the number of times the word dog is used in a story and retrieve facts about dogs (such as how many legs they have), but computers do not understand words the way humans do, and will not respond to the word dog the way humans do. The lack of real world knowledge is often revealed in software that attempts to interpret words and images. Language translation software programs are designed to convert sentences written or spoken in one language into equivalent sentences in another language. In the 1950s, a Georgetown–IBM team demonstrated the machine translation of 60 sentences from Russian to English using a 250-word vocabulary and six grammatical rules. The lead scientist predicted that, with a larger vocabulary and more rules, translation programs would be perfected in three to five years. Little did he know! He had far too much faith in computers. It has now been more than 60 years and, while translation software is impressive, it is far from perfect. The stumbling blocks are instructive. Humans translate passages by thinking about the content—what the author means—and then expressing that content in another language.

Intelligent or obedient?

Jeopardy! is a popular game show that, in various incarnations, has been on television for more than 50 years. The show is a test of general knowledge with the twist that the clues are answers and the contestants respond with questions that fit the answers. For example, the clue, “16th President of the United States,” would be answered correctly with “Who is Abraham Lincoln?” There are three contestants, and the first person to push his or her button is given the first chance to answer the question orally (with the exception of the Final Jeopardy clue, when all three contestants are given 30 seconds to write down their answers). In many ways, the show is ideally suited for computers because computers can store and retrieve vast amounts of information without error. (At a teen Jeopardy tournament, a boy lost the championship because he wrote “Who Is Annie Frank?” instead of “Who is Anne Frank.”A computer would not make such an error.) On the other hand, the clues are not always straightforward, and sometimes obscure. One clue was “Sink it and you’ve scratched.” It is difficult for a computer that is nothing more than an encyclopedia of facts to come up with the correct answer: “What is the cue ball?” Another challenging clue was, “When translated, the full name of this major league baseball team gets you a double redundancy.” (Answer: “What is the Los Angeles Angels?”) In 2005 a team of 15 IBM engineers set out to design a computer that could compete with the best Jeopardy players. They named it Watson, after IBM’s first CEO, Thomas J. Watson, who expanded IBM from 1,300 employees and less than $5 million in revenue in 1914 to 72,500 employees and $900 million in revenue when he died in 1956. The Watson program stored the equivalent of 200 million pages of information and could process the equivalent of a million books per second. Beyond its massive memory and processing speed, Watson can understand natural spoken language and use synthesized speech to communicate. Unlike search engines that provide a list of relevant documents or web sites, Watson was programmed to find specific answers to clues. Watson used hundreds of software programs to identify the keywords and phrases in a clue, match these to keywords and phrases in its massive data base, and then formulate possible responses.

The Kitchen Sink

Back in the 1980s, I talked to an economics professor who made forecasts for a large bank based on simple correlations like the one in Figure 1. If he wanted to forecast consumer spending, he made a scatter plot of income and spending and used a transparent ruler to draw a line that seemed to fit the data. If the scatter looked like Figure 1, then when income went up, he predicted that spending would go up. The problem with his simple scatter plots is that the world is not simple. Income affects spending, but so does wealth. What if this professor happened to draw his scatter plot using data from a historical period in which income rose (increasing spending) but the stock market crashed (reducing spending) and the wealth effect was more powerful than the income effect, so that spending declined, as in Figure 2? The professor’s scatter plot of spending and income will indicate that an increase in income reduces spending. Then, when he tries to forecast spending for a period when income and wealth both increase, his prediction of a decline in spending will be disastrously wrong. Multiple regression to the rescue. Multiple regression models have multiple explanatory variables. For example, a model of consumer spending might be: C = a + bY + cW where C is consumer spending, Y is household income, and W is wealth. The order in which the explanatory variables are listed does not matter. What does matter is which variables are included in the model and which are left out. A large part of the art of regression analysis is choosing explanatory variables that are important and ignoring those that are unimportant. The coefficient b measures the effect on spending of an increase in income, holding wealth constant, and c measures the effect on spending of an increase in wealth, holding income constant. The math for estimating these coefficients is complicated but the principle is simple: choose the estimates that give the best predictions of consumer spending for the data used to estimate the model. In Chapter 4, we saw that spurious correlations can appear when we compare variables like spending, income, and wealth that all tend to increase over time.

Introduction

The Democratic Party’s 2008 presidential nomination was supposed to be the inevitable coronation of Hillary Clinton. She was the most well-known candidate; had the most support from the party establishment, and had, by far, the most financial resources. Two big names (Al Gore and John Kerry) considered running, but decided they had no hope of defeating the Clinton machine. That left an unlikely assortment of lesser-knowns: a U.S. Representative from Ohio (Dennis Kucinich), the Governor of New Mexico (Bill Richardson), and several U.S. Senators: Joe Biden (Delaware), John Edwards (North Carolina), Chris Dodd (Connecticut), Mike Gravel (Alaska), and Barack Obama (Illinois). The nomination went off script. Obama was a first-term senator, a black man with an unhelpful name, but he excited voters. He raised enough money to be competitive in the Iowa caucuses and he persuaded Oprah Winfrey to campaign for him. Obama defeated Clinton by eight percentage points in Iowa and the race was on. Obama won the Democratic nomination and, then, the presidential election against Republican John McCain because the Obama campaign had a lot more going for it than Obama’s eloquence and charisma: Big Data. The Obama campaign tried to put every potential voter into its data base, along with hundreds of tidbits of personal information: age, gender, marital status, race, religion, address, occupation, income, car registrations, home value, donation history, magazine subscriptions, leisure activities, Facebook friends, and anything else they could find that seemed relevant. Some data were collected from public data bases, some from e-mail exchanges or campaign workers knocking on front doors. Some data were purchased from private data vendors. Layered on top were weekly telephone surveys of thousands of potential voters which not only gathered personal data, but also attempted to gauge each person’s likelihood of voting—and voting for Obama. These voter likelihoods were correlated statistically with personal characteristics and extrapolated to other potential voters based on their personal characteristics. The campaign’s computer software predicted how likely each person its data base was to vote and the probability that the vote would be for Obama. This data-driven model allowed the campaign to microtarget individuals through e-mails, snail mail, personal visits, and television ads asking for donations and votes.

Patterns in Randomness

I do an extra-sensory perception (ESP) experiment on the first day of my statistics classes. I show the students an ordinary coin— sometimes borrowed from a student—and flip the coin ten times. After each flip, I think about the outcome intently while the students try to read my mind. They write their guesses down, and I record the actual flips by circling H or T on a piece of paper that has been designed so that the students cannot tell from the location of my pencil which letter I am circling. Anyone who guesses all ten flips correctly wins a one-pound box of chocolates from a local gourmet chocolate store. If you want to try this at home, guess my ten coin flips in the stats class I taught in the spring of 2017. My brain waves may still be out there somewhere. Write your guesses down, and we’ll see how well you do. After ten flips, I ask the students to raise their hands and I begin revealing my flips. If a student misses, the hand goes down, Anyone with a hand up at the end wins the chocolates. I had a winner once, which is to be expected since more than a thousand students have played this game. I don’t believe in ESP, so the box of chocolates is not the point of this experiment. I offer the chocolates in order to persuade students to take the test seriously. My real intent is to demonstrate that most people, even bright college students, have a misperception about what coin flips and other random events look like. This misperception fuels our mistaken belief that data patterns uncovered by computers must be meaningful. Back in the 1930s, the Zenith Radio Corporation broadcast a series of weekly ESP experiments. A “sender” in the radio studio randomly chose a circle or square, analogous to flipping a fair coin, and visualized the shape, hoping that the image would reach listeners hundreds of miles away. After five random draws, listeners were encouraged to mail in their guesses. These experiments did not support the idea of ESP, but they did provide compelling evidence that people underestimate how frequently patterns appear in random data.

Bad Data

When I first started teaching economics in 1971, my wife’s grandfather (“Popsie”) knew that my Ph.D. thesis used Yale’s big computer to estimate an extremely complicated economic model. Popsie had bought and sold stocks successfully for decades. He even had his own desk at his broker’s office where he could trade gossip and stocks. Nonetheless, he wanted advice from a 21-year-old kid who had no money and had never bought a single share of stock in his life—me—because I worked with computers. “Ask the computer what it thinks of Schlumberger.” “Ask the computer what it thinks of GE.” This naive belief that computers are infallible has been around ever since the first computer was invented more than 100 years ago by Charles Babbage. While a teenager, the great French mathematician Blaise Pascal built a mechanical calculator called the Arithmetique to help his father, a French tax collector. The Arithmetique was a box with visible dials connected to gears hidden inside the box. Each dial had ten digits labeled 0 through 9.When the dial for the 1s column moved from 9 to 0, the dial for the 10s column moved up 1 notch; when the dial for the 10s column moved from9 to 0, the dial for the 100s column moved up 1 notch; and so on. The Aritmatique could do addition and subtraction, but the dials had to be turned by hand. Babbage realized that he could convert complex formulas into simple addition-and-subtraction calculations and automate the calculations, so that a mechanical computer would do the calculations perfectly every time, thereby eliminating human error. Babbage’s first design was called the Difference Engine, a steam powered behemoth made of brass and iron that was 8 feet tall, weighed 15 tons, had 25,000 parts. The Difference Engine could make calculations up to 20 decimals long and it could print formatted tables of results. After a decade tinkering with the design, Babbage began working on plans for a more powerful calculator he called the Analytical Engine. This design had more than 50,000 components, used perforated cards to input instructions and data, and could store up to one thousand 50-digit numbers.

We’re watching You

Humans often anthropomorphize by assuming that animals, trees, trains, and other non-human objects have human traits. In children’s stories and fairy tales, for example, pigs build houses that wolves blow down and foxes talk to gingerbread men. Think about these stories for a minute. The three little pigs have human characteristics reflected in the houses they build of straw, sticks, or bricks. The wolf uses various ruses to try to lure the pigs out of the brick house, but they outwit him and then put a cauldron of boiling water in the fireplace when they realize that the wolf is climbing up the roof in order to come down the chimney. The gingerbread man is baked by a childless woman, but then runs away from the woman, her husband, and others, taunting his pursuers by shouting, “Run, run as fast as you can! You can’t catch me. I’m the Gingerbread Man!” In some versions, a fox tricks the gingerbread man into riding on his head in order to cross a river and then eats him. In the version read to me when I was a child, a wily bobcat tries to lure the gingerbread man into his house for dinner, but birds in a nearby tree warn the gingerbread man that he is the dinner. The gingerbread man flees while the bobcat snarls, “Botheration!” The gingerbread man runs back home, where he is welcomed by his family and promises never to run away again. These are enduring fairy tales because we are so willing, indeed eager, to assume that animals (and even cookies) have human emotions, ideas, and motives. In the same way, we assume that computers have emotions, ideas, and motives. They don’t. Nonetheless, we are fascinated and terrified by apocalyptic science-fiction scenarios in which robots have become smarter than us—so smart that they decide they must eliminate the one thing that might disable them: humans. The success of movies such as Terminator and Matrix has convinced many that this is our future and it will be here soon. Even luminaries such as Stephen Hawking and Elon Musk have warned of robotic rebellions.

Beat the Market II

Nowadays, technical analysts are called quants. Being overly impressed by computers, we are overly impressed by quants using computers instead of pencils and graph paper. Quants do not think about whether the patterns they discover make sense. Their mantra is, “Just show me the data.” Indeed, many quants have PhDs in physics or mathematics and only the most rudimentary knowledge of economics or finance. That does not deter them. If anything, their ignorance encourages them to search for patterns in the most unlikely places. The logical conclusion of moving from technical analysts using pencils to quants using computers is to eliminate humans entirely. Just turn the technical analysis over to computers. A 2011 article in the wonderful technology magazine Wired was filled with awe and admiration for computerized stock trading systems. These black-box systems are called algorithmic traders (algos) because the computers decide to buy and sell using computer algorithms in place of human judgment. Humans write the algorithms that guide the computers but, after that, the computers are on their own. Some humans are dumbstruck. After Pepperdine University invested 10 percent of its portfolio in quant funds in 2016, the director of investments argued that, “Finding a company with good prospects makes sense, since we look for under valued things in our daily lives, but quant strategies have nothing to do with our lives.” He thinks that not having the wisdom and common sense acquired by being alive is an argument for computers. He is not alone. Black-box investment algorithms now account for nearly a third of all U.S. stock trades. Some of these systems track stock prices; others look at economic and noneconomic data and dissect news stories. They all look for patterns. A momentum algorithm might notice that when a particular stock trades at a higher price for five straight days, the price is usually higher on the sixth day. A mean-reversion algorithm might notice that when a stock trades at a higher price for eight straight days, the price is usually lower on the ninth day. A pairs-trading algorithm might notice that two stock prices usually move up and down together, suggesting an opportunity when one price moves up and the other doesn’t.

Beat the Market I

Back when scam artists sent snail-mail instead of e-mail, I received a letter that began “Dear friend,” a clear sign it was from someone trying to sell me something. Nonetheless, I read a bit more and saw this line highlighted in yellow: “IMAGINE turning $1,000 into $34,500 in less than one year!” Real friends don’t highlight their sentences, but I pushed on, thinking I might share this BS with my students. Sure enough, it was a con. The letter said that “no special background or education” was needed and that, “It’s an investment you can make with spare cash that you might ordinarily spend on lottery tickets or the race track.”Now I wasn’t sure that I wanted to share this letter, lest my students wonder where this company had gotten my name. I don’t buy lottery tickets or bet on horses. What had I done that made this company think I was a sucker? The letter claimed that, instead of wasting my money on lottery tickets and horse races, I could get rich buying low-priced stocks. For example, the price of LKA International had jumped from 2 cents a share to 69 cents a share in a few months, which would have turned $1,000 into $34,500. All I had to do was pay $39 for a special report that would give me access to “the carefully guarded territory of a few shrewd ‘inner circle’ investors.” The entire premise is ridiculous. If someone really knew how to turn $1,000 into $34,500, they would be doing it, instead of selling special reports for $39. Yet, we repeatedly fall for such scams because we are hard-wired to think that the world is governed by regular laws that we can discover and exploit. Stock prices cannot be random. There must be underlying patterns, like night and day, winter and summer. Our gullibility is aided and abetted by our greed—by the notion that it is easier to make money by buying and selling stocks than by being a butcher, baker, or candlestick maker. The inconvenient truth is that zigs and zags in stock prices are mostly random, yet transitory patterns can be found in random numbers. If we look for them, we will find them and be fooled by them.

Take Two Aspirin

IBM’s Watson got an avalanche of publicity when it won Jeopardy, but Watson is potentially far more valuable as a massive digital database for doctors, lawyers, and other professionals who can benefit from fast, accurate access to information. A doctor who suspects that a patient may have a certain disease can ask Watson to list the recognized symptoms. A doctor who notices several abnormalities in a patient, but isn’t confident about which diseases are associated with these symptoms, can ask Watson to list the possible diseases. A doctor who is convinced that a patient has a certain illness can ask Watson to list the recommended treatments. In each case, Watson can make multiple suggestions, with associated probabilities and hyperlinks to the medical records and journal articles that it relied on for its recommendations. Watson and other computerized medical data bases are valuable resources that take advantage of the power of computers to acquire, store, and retrieve information. There are caveats though. One is simply that a medical data base is not nearly as reliable as a Jeopardy data base. Artificial intelligence algorithms are very good at finding patterns in data, but they are very bad at assessing the reliability of the data and the plausibility of a statistical analysis. It could end tragically if a doctor entered a patient’s symptoms into a black-box data-mining program and was told what treatments to use, without any explanation for the diagnosis or prescription. Think for a moment about your reaction if your doctor said, I don’t know why you are ill, but my computer says, “Take these pills.” I don’t know why you are ill, but my computer recommends surgery. Any medical software that uses neural networks or data reduction programs, such as principal components and factor analysis, will be hard-pressed to provide an explanation for the diagnosis and prescribed treatment. Patients won’t know. Doctors won’t know. Even the software engineers who created the black-box system won’t know. Nobody knows. Watson and similar programs are great as a reference tool, but they are not a substitute for doctors because: (a) the medical literature is often wrong; and (b) these errors are compounded by the use of data-mining software.