The Robots Learn By Watching Us

Also golf, stamps, fake memos and Tesla conspiracies.

The Robots Learn By Watching Us

Machine learning.

There are two basic ways to do machine learning in investing:

  1. Program a computer to look at data about stocks to figure out what patterns, historically, have predicted the stocks that will go up.
  2. Program a computer to look at data about stocks to figure out what patterns, historically, have predicted the stocks that you will buy.

The advantage of approach 1 is obvious: Your goal is to buy stocks that go up. If you build a robot that buys stocks that go up, it will make you rich. If you build a robot that buys the stocks you would otherwise have bought, that robot might also make you rich, but only if you would otherwise have bought stocks that would go up. Even if you are good at buying stocks that will go up, you will sometimes buy stocks that go down. Even if the robot is good at doing whatever you program it to target, it will sometimes make mistakes. If the robot is trying to buy stocks that go up, there is one source of error (the robot). If the robot is trying to buy the stocks that you would buy, and you were trying to buy stocks that go up, there are two sources of error (the robot, you). If, as seems plausible, they compound, then approach 2 seems worse than approach 1.

But there are examples of approach 2. Steven A. Cohen’s Point72 Asset Management built a robot whose job was to figure out how Point72 made investments and then do that, “parsing troves of data from its portfolio managers and testing models that mimic their trades.” The target is not “stocks that went up” but rather “analyst recommendations”; the computer learns to recommend stocks that Point72’s human analysts would recommend, not the ones that an objective omniscient stock predictor would recommend. This approach has advantages too. For instance, it might create a more intuitive output. The worry with machine learning is often that it is a black box: The computer will tell you what stocks to buy, but not why, and you will be uneasy about whether the computer recommended a stock because it saw a pattern that you missed, or because it saw a pattern that isn’t really there, or because a squirrel chewed up its wires. But if the computer isn’t picking stocks that it likes, but stocks that it thinks you will like, then you are more likely to recognize what it sees in those stocks. The goal of investing is not just to get high returns but also to get understandable, explainable, replicable, robust returns, and to convince clients to invest with you. If you have a sense of why you are buying the stocks, that is good, even if maybe the stocks go up less.

The nice thing about stocks is that they either go up or down, and you have a lot of historical data about which did which. What if you want to build a machine learning program to hire employees? You can try approach 1: Look at data about job candidates and try to predict which attributes correlate with good employees. But this is much harder than it is with stocks. The data about candidates is messier and less complete. Measuring which employees are good is less objective and more multidimensional. 

And there is the important confounding factor that you only hire some of the candidates. You can backtest a stock-picking algorithm on all the stocks in the world, the ones you bought and the ones you didn’t, but you can only test an employee-goodness algorithm on the employees you hired. Perhaps the attribute that best predicts job performance is writing “Interests: Smoking marijuana and watching television” on a resume. If you never hire any people like that — out of anti-drug bias, or just random chance — then your machine-learning algorithm will never learn that they are the best. 

And so in practice a machine-learning hiring algorithm will either explicitly take approach 2 — “here is a stack of resumes, which are the resumes that look like people our hiring managers would hire?” — or will take approach 1 — “here is a stack of resumes, which are the resumes that look like people who do a good job at our company?” — but in a way that ends up looking a lot like approach 2. Either way you get something that recreates a lot of your fallible biased human choices in algorithmic form.

Here is a story about how Inc. built a machine-learning tool “to review job applicants’ resumes with the aim of mechanizing the search for top talent,” but then found out that it had a gender bias:

That is because Amazon’s computer models were trained to vet applicants by observing patterns in resumes submitted to the company over a 10-year period. Most came from men, a reflection of male dominance across the tech industry. 

In effect, Amazon’s system taught itself that male candidates were preferable. It penalized resumes that included the word “women’s,” as in “women’s chess club captain.” 

It is not clear exactly what data the algorithm is trained on, or exactly what it was targeting, but intuitively these takes seem plausible:

If you build a machine-learning robot and ask it “who are the best employees” and it goes out and examines the world and comes back and says “men, it’s men, men are the best,” then that tells you something surprising and complicated about either the world or your robot. But if you build a machine-learning robot and ask it “who are the candidates we like to hire” and it goes out and examines your hiring decisions and comes back and says “men, it’s men, you hire men,” then that tells you something quite straightforward about your hiring decisions.

Oh by the way:

Gender bias was not the only issue. Problems with the data that underpinned the models’ judgments meant that unqualified candidates were often recommended for all manner of jobs, the people said. With the technology returning results almost at random, Amazon shut down the project, they said.

“It hired good people but a disproportionate number of them were men” is a problem, but it’s a fairly common problem in the world. That more or less replicates a lot of regular human hiring processes. “It hired people totally at random” seems worse and weirder! Humans try not to do that! 

Elsewhere in robots.

Here is a roundup of “some of the ways technology is changing investment bankers’ jobs,” by automating investment banking tasks. Investment banking is a complex job that combines knowledge (about what potential targets or acquirers are out there, about what precedent deals looked like), analysis (about how much a company is worth, about how to structure a deal), and people skills (to get executives to hire you and to negotiate deals for them). You could easily imagine automating much of the knowledge and analysis, leaving investment bankers as just sort of charismatic well-dressed user interfaces for a computer that spits out what companies to buy and how much to pay. So:

Aryeh Bourkoff, previously a senior banker at UBS Group AG who now runs his own investment bank, LionTree LLC, has been working with Nobel laureate and cognitive behavior scientist Daniel Kahneman to develop a process of finding deal ideas to present to clients and eventually match clients with potential buyers or sellers. The program would be a tool for bankers, not a replacement for them.

So you meet with a big company’s chief executive officer, you play golf, you ask her about her kids, and then you propose that she buy a little private company that fits well with her business: It’s exactly the same as now, except that you got the idea for buying that private company from your computer algorithm, not from your previous rounds of golf with its CEO. You’ve outsourced the market-knowledge part of your job to a robot; you do the golf and small-talk part.

Still, remember the two possible robot approaches to stocks, or hiring. When you build a robot to generate deal ideas, and you ask it to generate ideas to pitch to a CEO, do you want it to generate deals that will be accretive for her company, or do you want it to generate deals that she will want to do? Obviously there is a ton of overlap between those targets, but they are not identical. CEOs have interests and desires other than shareholder value, and if you are a self-interested banker you might want your computer to optimize for deals that get you paid rather than for deals that are good. Perhaps the computer needs some people skills too.

The Stamp King.

This story about Bill Gross’s stamp collecting is an utter delight, mostly because everything about Bill Gross’s stamp collecting — one suspects, everything about Bill Gross’s life — is a reflection of Bill Gross’s bond investing. For instance he is apparently a top-down macro stamp investor:

“I thought it would be much like the art market, where there would be a provenance and you could go back in prior auctions and check it related to inflation and GDP growth and wealth creation,” he says. “As Charles [Shreve] told me, that approach was unique; nobody had ever thought about that in stamps.”

He collected data, from Bloomberg and elsewhere, on inflation, interest rates, and GDP growth, and used that annual rate of wealth creation to index against the few data points from “important” prior auctions, like the Ishikawa sale, going back to 1910, using vintage auction-house catalogs that themselves are collectible. He says he tried “to correlate it to what today’s price should be relative to what 1993’s price was.”

Adds Gross, “That similarity, that financial connection that I had with bonds, translated to stamps.”

I am obviously not much of a sophisticate, but I confess I didn’t realize that people thought that way about art. I had some vague notion that the value of a painting fluctuated with, like, changing artistic tastes, or the roster of particular billionaires who were drawn to a particular painting. But Gross’s top-down approach — which he brought from the art market to the stamp market — suggests that the global financial portfolio includes an allocation of X basis points to, like, Picasso, and the values of the Picasso paintings are determined principally by global wealth rather than by artistic criteria. In some ways: Duh. In other ways I am a bit disillusioned to see it spelled out so clearly. 

Also this is fun:

At first, he bought under the pseudonym Monte Carlo, named for his street in Laguna Beach, Calif. But he caught attention after spending more than $2 million at the 1993 sale of Ryohei Ishikawa’s collection — “that’s when he really stepped out and started spending a lot of money,” Shreve says. It “immediately impacted the stamp market in a very positive way.”

First: Bill Gross lives on Monte Carlo Drive, sure. Second: What does a “positive way” mean, there? (Shreve is Gross’s “curator.”) Of course his dealer thinks that high prices are good; it’s not so clear that the actual buyer should think that. But I suppose if you splash a lot of money and push up the price of stamps, then you have at least a paper mark-to-market profit in your early investing, which might make you more enthusiastic about continuing. It might also genuinely make the market more robust: Others will see prices going up and will pile in, so that when you do sell you can do so at a profit. (This is oddly reminiscent of Gross’s Total Return Exchange-Traded Fund at Pacific Investment Management Co., which was alleged to have bought a lot of odd-lot bonds in order to print strong early performance numbers and attract clients.)

Eventually though Gross wept for there were no more worlds to conquer, in stamp collecting:

Gross’ arrival in 1993 helped to reinvigorate stamp collecting, and now his abdication should give it new life. Gross stopped collecting because, as he puts it, he “filled up all the spaces.” By winning too completely, he ruined the game for himself.

He bought all the stamps, so he had to stop. It really is a little reminiscent of how he left Pimco, where his fund was so big that it had limited room to maneuver, and started over with a small fund at Janus. Maybe Gross will take the millions he made on stamps and go put, like, $1,000 into Beanie Babies just to make the game fun again.

Everything is … uh, whatever this is.

Sure okay:

Broadcom Inc., which is closing in on a purchase of CA Technologies Inc., said it is the victim of a fraudulent effort to raise national security concerns about the deal.

The company said a memo has been circulating among U.S. lawmakers that purports to be a Defense Department assessment, prepared for a U.S. national security panel, outlining the need for a review of the deal. The company said it was told by Defense Department officials that the memo is a forgery.

The Defense Department, in a statement, said: "Our initial assessment is that this is likely a fraudulent document."

So on the one hand, sure: You bet against the deal, you write a fake Defense Department memo opposing the deal, you send the memo to journalists hoping to get something published about it, the stock drops, you make money, you quietly walk away, I get it. But this is just kind of weirder, you know? Circulating your fake memo widely in Congress is not a typical element of a fake-press-release securities fraud. “It’s unclear whether parties may have been trying to derail the deal or plant the memo in an effort to profit from share price moves,” reports Bloomberg. What if they derail the deal? What if congresspeople read the memo and are like “the Defense Department makes some good points” and then the Defense Department is like “it wasn’t us” and the congresspeople are like “ehh I don’t know what to believe anymore, let’s block this deal”?

I guess my point here is that, as a society, our grip on reality seems pretty tenuous. Ordinarily if you want to fraudulently convince investors that a deal is in danger, the simplest way is to publish a fake document to investors saying the deal is in danger, because investors are, typically, relatively easy to trick. But in the U.S. in 2018, if you want to fraudulently convince investors that a deal is in danger, maybe the simplest way is to write a fake document and give it to Congress and actually cause Congress to endanger the deal? Maybe infecting reality with fraud is now easier than infecting financial markets with fraud?

Tesla Short Seller Criminality Watch 2018.

People on Twitter — including, apparently, prominent short seller Jim Chanos — have been enjoying this article by a bullish Tesla Inc. investor titled “A Field Guide To Potential Securities Violations By Tesla’s Foes — In Depth.” “Without brilliant work like this I never would have known that every time Tesla stock goes down a crime has been committed,” tweeted E.W. Niedermeyer. That is definitely the tone! A sample paragraph:

Some of the patterns of stock price manipulation we see all the time with Tesla are so common they have received names, such as mandatory morning dipdip on steroidssticky diplow-volume dip into close, and capping.

There are people in the world who claim to dislike the passive voice, but maybe they have just never seen a passive construction as good as “they have received names.”

The claims here are long and varied and mostly without evidence, but the gist is that Tesla short sellers manipulate the stock downward by (1) selling stock to push the price down to drown out the effect of good news, (2) insider trading on advance knowledge of negative news articles and research reports, and (3) controlling the media. One oddity is that the author seems to think that every time Tesla goes down — or even doesn’t go up — that is caused by new short selling. But in fact, according to Nasdaq data, Tesla short interest peaked at about 39 million shares in April, and ended September at about 33.6 million shares, meaning that short sellers have been net buyers of more than 5 million shares in recent months. If Tesla shorts are constantly selling stock to drive down or cap its price, then they are also constantly buying back even more stock, presumably pushing up the price.

So I mean it’s probably wrong. But who cares? Isn’t it sort of enchanting? Isn’t it nice to think that everything bad that happens to you is the work of a deep conspiracy of well-funded shadowy enemies who can move the market and control the media? Isn’t that better than just thinking that, you know, you bought stock in a money-losing car company at a $50 billion valuation and then its chief executive officer started repeatedly shooting himself in the foot?

Oh oh oh also I can actually clear one thing up for him. He notes that someone bought 1,000 out-of-the-money Tesla puts on Thursday, Sept. 27, expiring on Friday, Sept. 28, and made a quick $1.25 million:

September 27 was the Thursday in which the SEC disclosed after market close that it was pursuing a suit against Elon Musk. TSLA closed in the mid 260s the next day, Friday, when the puts expired. Buying nearly a thousand far-out-of-the-money puts that expired the next day suggests either knowledge of a big negative event that was coming or a suicidal trading methodology. Tesla would have to drop 4% in a day just to get the bet close to the break-even point. It’s also very concerning that the apparent insider trading came from a tip about an upcoming SEC action, of all things. Just how did the buyer of the puts find out?

We’ve discussed this! On Wednesday, Sept. 26, I wrote that “if Tesla Inc. announced tomorrow that Elon Musk stole all its money and fled on a rocket to Mars, it probably wouldn’t even be front-page news.” And someone read that, took it … somehow … seriously, and went and bought some short-dated out-of-the-money put options on Tesla stock. And then it happened to play out weirdly close to what I suggested. I mean, I know of one person who claims to have bought a few puts on my (joking) advice, but for all we know all 1,000 puts were similarly bought by people taking Money Stuff too seriously. Perhaps Money Stuff is secretly, and accidentally, controlling the anti-Tesla conspiracy. 

“Crypto is the Mother of All Scams and (Now Busted) Bubbles While Blockchain Is The Most Over-Hyped Technology Ever, No Better than a Spreadsheet/Database.”

That’s the title of Nouriel Roubini’s testimony to Congress about crypto today. It does seem to cover his main points succinctly.

Things happen.

Not Just ‘Loco’ Fed: Why Equities Are Suddenly Selling Off Now. Bitcoin Tumbles as Cryptocurrencies Join Global Asset Selloff. A $1 Trillion Powder Keg Threatens the Corporate Bond Market. Lampert Won’t Give Sears Another Lifeline. James Murdoch in line to replace Elon Musk as Tesla chair. Inside the C.E.O.’s Social Media Meltdown at Deciem. Newsweek’s Ex-Parent Company Charged With Defrauding Lenders. What Happens When Regulatory Capital Is Marked to Market? Najib Razak unrepentant over 1MDB scandal. An interview with Sugata Ray, author of that hedge-fund-managers’-cars paper. (Earlier.) A profile of Warren Mosler, a founder of Modern Monetary Theory. Retirement forests. Fat bears. A Sommelier Scandal Is Rocking the Wine World. Flight Delayed After Woman Brings ‘Emotional Support Squirrel’ on Plane. 

If you'd like to get Money Stuff in handy email form, right in your inbox, please subscribe at this link. Thanks!

Like Money Stuff? Subscribe to Bloomberg All Access and get much, much more. You'll receive our unmatched global news coverage and two in-depth daily newsletters, The Bloomberg Open and The Bloomberg Close.

    This column does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.

    To contact the author of this story:
    Matt Levine at

    To contact the editor responsible for this story:
    James Greiff at

    Comments 0