The entire big data movement has its roots in the research of that industry-leading cohort at IBM, which took advantage of the company’s vast troves of speech examples and world-class computer power to recognize human language better than anyone else. Applying the Markov tools to money and investing, the Renaissance team saw that if you can predict the next word in a sentence, you can predict the likely next price for stocks, commodities, or currencies. With clusters of supercomputers running at sufficient velocity, you could beat every short-term market you could access and measure. In 2009 Simons retired and named Mercer and Brown co-CEOs of his company.
Mercer’s IBM boss, Fred Jelinek, was a protégé of the MIT information theorist Robert Fano and a student of Claude Shannon. He saw speech recognition as an information theory problem—an acoustic signal and a noisy channel. Citing the content-neutral concept behind his speech-recognition successes, Jelinek proudly declared, “Every time I fire a linguist, the performance improves.” Renaissance’s approach similarly spurns any direct kibitzing from fundamental analysts or anyone who knows anything special about particular companies.
Relying on its world-leading complement of mathematicians and physicists, Renaissance “avoids hiring anyone with even the slightest whiff of Wall Street bona fides,” comments James Owen Weatherall in The Physics of Wall Street. Instead, it takes in vast troves of information from analyst reports, government reports, newspaper stories and newswires, in addition to prices and trades wherever it can find them. All this material, produced by human effort and brain power, enables the Markovian system to ignore human intentions and purposes.
When I wrote my book Microcosm in 1989, I had come to appreciate the amazing accomplishments of the IBM team.7 But it was not until 2016 that I managed to wangle an invitation to interview Mercer. My goal was to find out whether he had found the secret of Midas or had merely learned the hapless king’s “lesson.” Granted a wish to turn anything he touched into gold, Midas made the mistake of giving his beloved daughter a hug.
Driving down Long Island and looking for Mercer’s home in Head of the Harbor, I found a turnoff from 25A that aimed me down a long dirt road with sand on all sides under a green bower. Twisting left and right through a state park, avoiding hikers and bicyclists amid gusts of dust, I drove for a quarter of an hour before arriving at Mercer’s gate. There I announced myself through a microphone on a post. “Enter the gate,” I was told, “and then drive—very slowly—up the driveway to the house.”
I did as I was told and parked near a three-story structure of classical design overlooking Stony Brook Harbor off Long Island Sound. This in Markov terms is an “absorbing state” (no further turns). You have arrived. In my investigation of computers, information theory, Markov, and money, I imagined that I had penetrated to the secret heart of Google’s intellectual regime across the continent in Silicon Valley.
I was steered into a living room decorated with full-length portraits of Mercer’s daughters, Heather Sue, Rebekah, and Jenji. A mathematician and intellectual leader, Rebekah represents Mercer on the boards of such conservative think tanks as the Heritage Foundation and the Manhattan Institute.
The portraits easily held my attention until Mercer arrived, a handsome, self-effacing man in a gray suit with short-cropped gray hair. He was no-nonsense. After a minute or so of pleasantries, we plunged into a discussion of his investment strategy, beginning with a question about the velocity of Renaissance transactions and supercomputers.
“Velocity,” Mercer told me, “is not necessarily positive. It can be created by transactions that clearly in no way benefit the economy. For example, I could buy a car from someone for a thousand dollars and then sell it back to him for the same thousand dollars. To the econometrician, this would look as if two cars had been purchased when in fact nothing had changed at all. . . . ” And please, call me Bob.
Focusing my mind on velocity’s contribution to the Midas enigma was the work of Jaron Lanier, the shaggy sage who both invented virtual reality and identified the Siren Server. Lanier writes, “Siren Servers are usually gigantic facilities, located in obscure places where they have their own power plants and some special hookup to nature, such as a remote river that allows them to cool a fantastic amount of waste heat.”8 It did not seem to apply to the Renaissance datacenter in Long Island, but I immediately thought of Urs Hölzle’s Google facility in The Dalles by the Columbia River.
“This new class of ultrainfluential computers comes in many costumes,” writes Lanier. “Some run financial schemes such as high-frequency trading, and others run insurance companies. Some run elections, and others run giant online stores. Some run social network or search services, while others run national intelligence services. The differences are only skin deep. . . . ”9
“A Siren Server is a powerful computational resource that out-computes everyone else on the network and seems to grant its owners a guaranteed path to unbounded success at first”—thus its siren appeal. “But the benefits are illusory,” warns Lanier, “and lead to a grand failure before long.”10
I imagined that Google would ultimately meet this fate. But Mercer and his colleagues at Renaissance Technologies had apparently evaded the fate of Midas. There was no sign of anyone starving on a pile of gold in the midst of a wasteland.
Under the guidance of the cerebral cybernetic team of Mercer and Brown, Renaissance’s Medallion Fund has reportedly averaged a yield of roughly 40 percent every year, through up and down markets, for close to twenty years. Mercer and his consort of superstar scholars have, mutatis mutandis, excelled everyone else in the history of finance. Though Mercer is famous for his political role funding Republicans (Simons and Brown finance Democrats), he and his associates remain obdurately obscure with their unique achievements, hidden Markov chains of gold.
Unlike its West Coast counterpart Google, the Renaissance group completely escaped the perils of the Great Recession, which humbled so many hedge funds and big banks. During the crash of 2008, after extracting the industry’s highest fees—a vertiginous 5 percent of money under management and 44 percent of the profits—the Medallion Fund was said to be up 80 percent. Other hedge funds were down an average of 17 percent, and the S&P was down 40 percent.
The next year Medallion scored profits of more than a billion dollars and was ranked number one among all hedge funds. Mercer points out that my numbers are inconsistent. I grant his point. Take them as rough estimates, concocted by financial journalists in the face of an obsessively secretive industry. They roughly tell the story of a company on a Markovian ride of awesome dimensions.
With more than $65 billion currently under management, Mercer’s team relies on racks of Renaissance workstations linked to form supercomputers. They parse immense Markov chains of ordered data to find filigree “ghosts” of tradable correlation. Like Google’s PageRank and its Deep Learning successes with language translation and games, like IBM’s earlier speech-recognition breakthroughs, and like “Watson,” IBM’s supercomputing master of Jeopardy searches and chess strategies, it is founded on ever-faster processing of pure statistics from ever-larger databases.
As James Simons explained in a speech in 1999, “Efficient market theory is correct in that there are no gross inefficiencies. But we look at anomalies that may be small in size and brief in time. . . . We’re always in and out and out and in. So we’re dependent on [intense] activity to make money.”11 Their strategy is based on round-the-clock processing of terabytes of data in search of correlations that yield profit opportunities. “Some of the signals that we have been trading on without interruption for fifteen years make no ‘sense.’ Otherwise someone else would have found them,” Mercer acknowledges. “But there is no question from a statistical point of view that they work.”
As I mentioned to Mercer, I have often expressed my disdain for this approach as an “outsider trading” scandal. If the investors do not understand the reasons for their success or provide significant original analysis, they
do not increase the knowledge that underlies all productive investment in capitalism.
The Renaissance method seems to violate the Turing-Gödel principle that all logical systems need “oracles,” sources and assumptions outside themselves. A logical scheme or computer program that merely finds patterns in vast batches of data will eventually be governed by its environment, of which it is a creature. Predicting the future, it is ensnared in the entropy of its past—the chain of observables and their hidden derivatives. It cannot anticipate the human creativity that propels all progress.
As Caltech’s Carver Mead has said, “the only adequate model of the galaxy is the galaxy.” The database can grow galactically, but it cannot substitute for the patient acquisition of specific and singular information about business plans, inventions, and technologies deep within companies.
Mercer responds: “The fact is, we are the oracle with respect to problems involving the question, ‘What does the history of past market reactions in the presence of then-known information about the then-known current state of the market tell us about the future of the market?’ We achieve this status [as oracle] by putting more brain power and more computing power into it than anyone else does.”
Taking both long and short positions, the funds are market neutral. Without understanding the actual situation, Markovian tools can succeed whether the market booms or crashes. Hence the amazing performance in 2007 and 2008. Without relying on the outsized leverage that brought down other funds, Renaissance thrives by processing more data, building larger Markov chains, ferreting out more correlations and probabilities, and executing more trades than anyone else.
A venture capitalist on Sand Hill Road in Palo Alto investing in an embryonic Google after acquiring intimate knowledge of its technology may make a thousand-fold return over five to seven years. A firm such as Renaissance might make a thousand trades in a day harvesting the tiniest anomalies. With modest leverage and relentless twenty-four-hour trading around the globe, Medallion could make far more money than a venture capitalist without knowing any details about the technologies and business plans behind the equities, currencies, or securities traded. This is the financial counterpart of Markov models at Google translating languages with no knowledge of them.
Believing as I do in the centrality of knowledge and learning in capitalism, I found this fact of life and leverage absurd. If no new knowledge was generated, no real wealth was created. As Peter Drucker said, “It is less important to do things right than to do the right things.” Effectiveness is more important than efficiency. Renaissance’s improvement in market efficiency is small compared to the yield. As a result, too much American capital is migrating to Siren Servers and shunning “Zero to One” creative investments. Computerized index funds that “buy the market” thrive while IPOs languish. No net wealth is created, but money is arbitrarily siphoned off and redistributed in a zero-sum game.
I pointed out that Renaissance’s “neutral” approach benefited from the counterproductive futility of insider-trading rules and fair disclosure requirements, which hobble rivals who are using their human brains in real time. Mercer disarmingly agreed. By remaining obtusely obsessed with the usually innocuous edge gained by people who laboriously investigate inside of companies, the Securities and Exchange Commission has driven the vast bulk of trading to purely algorithmic forms. You can’t indict a computer. But you can’t make a creative investment with one, either.
I prefer to believe a non-Markov model will win. Why? Newton’s insight that the “white light is a mixture, but the colored beams are pure” inspired Jean-Baptiste Joseph Fourier, a century later, to use Newton’s infinite mathematical series to describe the effect of prisms.12 The “Fourier transform” can be used not only on light rays but on any time-based signal—sound waves, for example—that we want to break down into its constituent cycles. With a formula that now pervades wireless telephony, acoustics, and optics, Fourier showed that any complex wave—from heat waves to operatic solos to WiFi signals to economic or monetary cycles—could be expressed as an overlapping series of regular sine waves like pure sounds or colors.
In finance, a Fourier model would move from the time domain of the record of trades, one after another in a Markov chain, to the frequency domain depicting the pure frequency components of the trading pattern. Converting from the time domain of all Medallion’s trades, for example, we could discover an underlying set of pure frequencies that combines information about the amplitude and the power of each investment.
Since the power of a wave rises with the square of its amplitude, large and long investments would be exponentially more significant than a series of small trades. Wavelets would be exponentially less potent than tsunamis. “That is why ‘Flash Boys’ do not finally make much money,” notes Mercer. The vast enterprise of Renaissance, selecting and collecting data and refining algorithms of discovery, he suggests, goes far beyond mere high-velocity trading.
To the frequency data, my model adds profits, the economic manifestation of entropy—the unexpected dimension of returns beyond the interest rate, which reflects average and predictable returns. Derived from Claude Shannon’s information theory, entropy in my model is surprise. Small and temporary anomalies are unsurprising and low-entropy.
Correcting for leverage, I contend that profits that merely reflect borrowing power do not usually contribute to the learning process. They reveal willingness to accept a level of calculable risk rather than singularities of creative learning. Such profits are predictable and thus low-entropy.
The Siren Server search for momentary correlations falls within the ambit of Stanford Nobel physicist Robert Laughlin’s critique of the science of frothy phase changes. Parsing the chaotic ebullition of water as it comes to a boil, for example, is a fool’s errand called “chaos theory.”
From its earliest days (when it was called Monemetrics), Renaissance has been active in foreign exchange markets. The very essence of froth on world markets, currency trading is roughly one hundred times greater than all the world’s stock market transactions and twenty-six times more voluminous than the world’s GDP. The oceanic currency markets are full of Laughlin froth to be parsed by computers for short-term anomalies. Even at Renaissance’s modest level of leverage—reported to be five to one—these trades may produce massive profits. But the profits do not contribute to the processes of entropic learning that constitutes all economic growth in an economy of knowledge.
In his defense, Mercer appeals to the vital role of money markets and banks’ aggregation of available wealth in the rise of the British Empire, citing Walter Bagehot’s Lombard Street (1873). There is a difference, however, between nineteenth-century London and today.
Bagehot’s Britain operated under Newton’s gold standard and system of the world. The currencies that central banks manage today have no anchor in gold and thus suffer from the self-referential circularity of all logical systems not moored to reality outside of themselves. In the United States, unmoored Markovian money can be manipulated at will by the Federal Reserve in the interests of its sponsors in government and their pseudo-private cronies.
Unmoored money changes the culture of capitalism. Wall Street banks relish volatile currencies, their downside protected by government. Main Street and Silicon Valley want stable money for long-term investments, the upside guaranteed by the rule of law. The governments of the world, their money unmoored, favor finance over enterprise, shortening the time horizons of economic activity. Among fast traders the rhythms are reduced to seconds and the economy endures a hypertrophy of short-term finance.
Mercer’s two careers illustrate the difference between entrepreneurial creativity and “market neutral” financial strategies. Market-neutral trading is a Midas touch in the financial system. It consists mostly of zero-sum maneuvers and has little engagement with the saga of creative human progress. It improves the efficiency and liquidity of markets at the cost of creating Siren Servers that lure the unwary into steril
e fields of algorithmic finance.
At IBM, by contrast, Mercer and his colleagues under Jelinek achieved a permanent advance in computer science, information theory, and speech recognition. Their discoveries are behind the Siri system in your iPhone, hands-free calling in your car, and the growing success of machine translation. They enabled the ever-improving responsiveness of voice interfaces to the cloud computing technologies in the new generation of Internet progress.
In the process, Mercer and his team pioneered the field of big data, which dominates the current computer paradigm. Competing with Kurzweil and other pioneers of AI-based systems attempting to duplicate human experts—from chess to translation—the IBM team faced the possibility of refutation and failure. Their advances, therefore, exhibited the Popperian power of falsifiable knowledge, the source of all new wealth under capitalism.
Big data today has become the system of the world of the Google era. But Lanier offers a portentous warning. “Your superior calculation ability allows you to choose the least risky options for yourself, leaving riskier options for everyone else. . . . ”13 He observes that “networked finance kept on pretending it could eject risk out into the economy at large, like a computer radiating waste heat with a fan, but it became as big as the system. [In 2008 and 2009, the] computer melted.”14
I left Head of the Harbor grateful for Mercer’s time and astounded by his achievement, which I believe is as impressive as the achievement of Google. But I concluded that this system of the world is obsolescent. It is based on big data that will face diminishing returns. It is founded on frequencies of trading that fail to correspond to any real economic activity. It feeds on mathematics of randomness that blur the differences between value creation and noise generation. Its source in memoryless Markov processes ultimately will bring the model back to the inevitable failures of gambler’s ruin.
Life After Google Page 9
