D. E. Shaw & Co.

D. E. Shaw & Co. - 1992 to 1994 Quant - "rocket scientist"

I started working for DESHAW in Manhattan, NYC, in early 1992. When I interviewed there, I met with Charles, the head of recruiting. I recall Charles dropping a comment that my resume seemed more like a 60-year-old than a 25-year-old, given the depth and breadth of my experiences. My resume at the time was about 10 pages and included in depth descriptions of my projects at UVa, Myarc, RPI, Chase, Electro Fiber Optic, and other personal projects. He asked my SAT score, for which I mentioned the National Merit Finalist status and a 1580 SAT score from my junior year in high school. After passing his initial vetting, I then interviewed with maybe half a dozen team members, many of whom were "department heads" who were hiring for their departments. Anne for ML forecasting and optimization algorithms, using multidimensional calculus based statistics, David Seigel for computer infrastructure, John Overdeck who was a Math Olympiad medal winner, Lou who I believe managed Anne and was also setting up a department to specialize in derivatives, Christos Zoulas who was inducted in the Unix hall of fame for NetBSD contributions, and Jeff Bezos who was leading an initiative to build a "third market" trading operation that would be a market-maker in over 2000 separate securities. After meeting with each of those people, I met with David Shaw, the owner of the company who asked me which of the people I had met I wanted most to work with. My reply was to work with Jeff Bezos and create the third market trading operation.

In my first visit to their office, which was in Tower 45 on 47th just off of Times Square, I could smell the new paint from their recent office move from above the bookstore, as well as the glue from carpet layers who were still actively installing more carpet. One quirky thing that still stands out in my memory is the original coffee maker in the kitchenette off of the main lobby -- it was not connected to a sink drain of any sort, even though the front of it had an overflow drain, and someone had put a sign on the front of the coffee maker, with an arrow to its drain, stating "this IS NOT a DRAIN -Magritte" Many of the offices on those top two floors of the building they rented were still empty. The elevator lobby and the main entrance hall were two stories tall, and the far wall from the elevators, overlooking 47th st, had a balcony connecting the east and west wings of the second floor. If I recall correctly, I was employee number 24, and we grew to over 100 people in the next two years.

I came to learn that all of the first 24 employees were supergenius in their own right and that's the only people Charles was interested in hiring -- intelligence over experience. We were a group of "rocket science quants" who employed two registered traders to leverage their license to trade legally. The computer and the algorithm would propose all of the trading decisions, and the registered trader's only job was to review the proposed transactions and hit the enter key. The NASD had a regulation at the time that there had to be a human in the loop, approving each trade. We sometimes joked that we could put a chicken in the loop and train it to peck the enter key whenever the computer proposed a trade. However, having a human in the loop allowed us to learn valuable feedback from time to time -- such as knowledge from our floor trader which had not made any news ticker such as Bloomberg yet, and those two registered traders were allowed to use their own discretion and reject proposed trades. Such discretion arose a few times a week, and when the traders broke or rejected the computer trades, we generally debriefed as a learning exercise to see if we could update the predictive algorithms to learn about such events and make adjustments before the trader rejected transactions. Due to our algorithmic automation, our trading volume on most days was tens of millions of shares, and on many days we traded over 80% of the NASDAQ total volume, at times nearing a billion shares per day (in 1992, that was nearly unheard of.) In 1992, our arbitrage model only had about a dozen forecasting criteria, but that grew to 20 or so by the time I left DESHAW.

Jeff's office was the far northeast corner office of the top floor, at the end of a hallway which had perhaps ten offices total. We had no intrinsic business interest in the third market making operation other than to provide another critical signal to the arbitrage algorithm, which was to provide an estimate of the depth and momentum of the "specialist" books from the NYSE trading floor. Those specialist books were generally secret from the public unless the specialist took a personal interest in you as a floor trader. The book, as it were, represented both the visible bid/ask, with their prices and available shares, as well as hidden latent demand at prices outside of the published bid/ask. In 1992, even the limit order books were not shared publicly, and were not available as a trading signal for algorithms such as ours. The general hypothesis, which I did not fully agree with, is that if we build a market-making capacity parallel to NYSE, NASDAQ, and other exchanges, that our internal book for each stock would come to resemble the book of the primary market-maker on NYSE or NASDAQ. We then intended to use our internal market making books to predict buy/sell demand for each stock which we became a market maker in, and use those demand predictions as input to our algorithmic trading model.

We acquired a "Stratus" computer which ran an operating system built using the Pascal language, and acquired market making software from a vendor who had written their software for the Mexican stock exchange. We also hired Simon who had been one of the primary developers of that software. We hired a former president of the NASD to leverage his network in our sales efforts, since in order to obtain order flow in that day and age, we had to place our proprietary terminal in key brokerage offices around the nation so that somebody in that office could key in their orders. (No networked SaaS at the time) My initial job, given my experience with creating GUI online banking for Chase, was to create a GUI terminal for our order entry system. Our terminal also allowed users to create "baskets" or small portfolios of orders which were "all or none".

I did one field trip to Boston to assist a brokerage there to bring up our terminal -- and I travelled with a portable, battery powered, 300 baud cellular modem which was slightly larger, and much heavier than, a rectangular lunch box. More like an ammo can. That setup was used to bring up and verify our trading software at many other brokerages in New York. I also made trips to Chicago and San Francisco to set up brokerages in those locations. The development of the GUI software, and the onsite visits to customers took up most of 1992 and the start of 1993. At some point in early 1993 the Stratus software crashed. It was not even a core dump, but a solid hang, and given our SLA of 10-15 minutes during the trading day, the entire system was rebooted without a core dump or any definitive evidence other than incomplete log files. I sure felt like I had seen this before, and put my hypothesis to the team that there was a mismatched mutex lock, with a system deadlock as a result. Nobody actually believed me, especially the Stratus support engineers who were confident that they had been shipping for years with no known deadlocks. In any event, I managed to get a copy of the VOS source code from them, and initially searched the text of the source for all locking behavior -- I found a few suspects wherein error handling code did not release locks that were held in at least 3 subroutines, including one which completely matched the evidence we had from the crash. Even with my findings, the Stratus support engineers denied any problem. We had one more similar crash two weeks later, and a month after the initial crash, we received a system update for VOS, which remarkably had the mutex fixes I had advocated for the error handling paths. Imagine that.

In early 1993, David was appointed to President Clinton's "Council of Economic Advisors", and our office started getting regular visits from George Stephanopolous several times per month.

At one point I learned we were paying a lot of money to Lotus One Source for the historical market data used for training the algorithm. The market data was packaged on CDROMS, and if I recall correctly, was $10,000 per month of historical data. When I learned that, I advocated for creating a pool of mass storage wherein we would bulk log the real time feeds we were already paying for, and then mine the logged data for training of the algorithm. I made the business case using the same metrics we used at HP -- BET, ROI, and ARR -- hard drives to store log data were much cheaper than the Lotus subscription, and also had more metadata and timing data than Lotus -- even in a fault-tolerant RAID configuration.

Sometime in December of 1993, our Third Market operation was receiving enough order flow to start building a meaningful book of trades across all 2000 symbols we were making markets in, we were ready to scale and add more customers. We were also confident in the signal we were providing to the arbitrage algorithm ... however, our margins were negative, and we were losing about $0.25 per trade -- with that in mind, David mentioned it would be disastrous to ramp volume, or "lose money at scale", especially since we had not finished measuring the impact of the new signal on potential profit from the algorithm itself.

One interesting thing about our arbitrage algorithm is that we could literally see, in the market data, when another company launched a high volume algorithm similar to ours. The only one we knew of for sure was Madoff, and in early 1994, we had observed signals due to at least one other player entering the market.

I had also become very interested in the arbitrage opportunities I saw in currencies, especially given the leverage banks would offer on such portfolios.

Over the course of 1992, 1993, and early 1994 I was invited to many get-togethers at Jeff's apartment on the upper west side ... if I recall correctly it was on 87th, two and one-third blocks from Central Park West, on the north side of the street. During one such event, Jeff shared that one of his passions, if he ever had the resources, was to build his own space program -- I was delighted, years later, when he followed through on that dream to launch Blue Origin. On another occasion, I was at his house with about 8 other people when he and MacKenzie announced their engagement -- and how she was the one who initiated dating with him due to his shyness at the time. As for their wedding, to the best of my knowledge and recollection they were not married before I left DESHAW -- my recollection contradicts some public timelines which suggest that they were married in 1993.

Another factor in our research is that many of us kept current with academic research papers that identified "market inefficiencies" -- and we would call and email the academic authors of those papers who were often super excited that anyone had read and understood their papers. They would let us drill into the details of their assumptions and their research data sets, and we would ask them what sort of other anomalies they had observed, often citing to work of other academics who were publishing papers at that time. For the most part, those papers ended up with the conclusion that the strategy and findings could not be profitable because the inefficiencies were smaller than the cost of trading -- not realizing, as we had, that we could negotiate our cost of trading far below the retail market rates cited by the papers because our portfolio allowed our clearing firm (Bear Stearns) to short other customer's positions against our long positions as they managed their own market exposure. For example, our clearing cost was under $0.01 due to our negotiations with Bear Stearns, a research paper would propose an inefficiency of $0.04, but then back off because the retail cost of settlement was over $0.05, and therefore "not profitable" in the eyes of academia.

I never really subscribed to the theory that "removing market inefficiencies", which is what our arbitrage algorithms did, would benefit anyone in society other than our firm. David and other senior members such as Peter and Lou would wax poetic about how our correction of those inefficiencies benefited the general public and promoted the social good. It was how they stroked their own egos, but did not feel right to me. Our algorithm thrived on: (1) volatility, not on long term stability or growth, and (2) our ability to constantly snipe between the bid and the ask to move large volumes of stock, usually without affecting the spread seen by retail investors. In essence, the sniping was an irritation to the specialist, who crossed our trades in order to maintain their spread.

New Years Eve 1993-1994, a Friday, I spent in the conference room of DESHAW where we had a great view of the "ball drop" in Times Square. There were about 10 people in the office that evening, and we started talking about the Internet ands its future. We had some ideas which were exciting enough that we got David to call an all-hands meeting later the next week in that same conference room to brainstorm ideas. On a dreary afternoon, Thursday Jan 6 1994, perhaps 60 people attended (we had added another 2 or 3 floors of the building in expanding our office space since I joined). In the meeting, Charles brought up the idea of selling books on the Internet, a virtual bookstore, and a number of other ideas were bounced around. I had shared an idea of an expert system that would help consumers make product decisions, and once the right fit product was identified, create an auction where providers of goods would bid to get the business of the consumer (somewhat backwards from the EBAY marketplace created later.)

The week after the all-hands, I set up a meeting with David to discuss my ideas and the general concept of profit sharing and equity going forward. There were a handful of us who helped make a $50 million windfall on some derivative trades during 1993, and part of my discussion with David was how that would benefit any of the team that advocated for and made the trade. David was not receptive to my ideas about profit sharing or equity at that time. He made it very clear that it was his relationship with the investors, and he was the Managing Partner with full liability if we lost money (which the firm never had lost up through that time) and that the base salaries we were receiving were more than sufficient compensation for the profits we made. After payments to investors and salaries, the remaining profit was distributed something like 97% to David, 1% to Peter (the technical cofounder who first created and built the entire arbitrage system), and 2% to the remaining 100 people in the company. The distribution did not work for me, and I told David that I would complete some projects and likely leave the company before summer unless he had a change of heart. I also asked permission to keep the rights to my shopping idea, and my currency trading ideas, to which he agreed. He even offered to invest in the currency trading idea with his own money if and when I launched that.

My first stop after the meeting with David was with Jeff Bezos, my immediate manager -- I told him I would be leaving soon, but that I would help transfer my tribal knowledge to some of the new people, and I shared the complete conversation with David regarding equity and profit sharing ... Jeff was surprised, and I challenged him to go ask David the same questions, to see if he would get the same answers, and if he was happy with those answers. Jeff did so, and he was not happy either.

I interviewed with Apple Computer the very next week, starting the exodus that included Jeff, and other team members at DESHAW I had worked closely with. Once word got around about my imminent departure, I got referred to another firm for which I did take an interview -- named the "Princeton Group" or something similar -- their office was just a few miles north of the Princeton University campus, and they were creating a fund and infrastructure to do currency arbitrage. I was excited to see their leverage model (often 50:1 on long positions), but ultimately choose Apple because financial arbitrage "had no soul", no redeeming qualities for society from my perspective. I left for Apple in April, Jeff left in May, with MacKenzie -- to start Amazon, with the bookstore idea we had discussed on that dreary Jan 6th -- Jeff had also asked David for permission to use that idea, and David said yes, much to his later chagrin. Others followed that summer. All-in-all, the exodus I led resulted in the departure of 7 of the original 24 people in the firm.

Things that happened after I left:

1. David had a change of heart about equity and startup investments, probably after seeing Amazon's successful IPO.
2. DESHAW spun off Juno Communications which became a leading nationwide dial-up ISP,
3. DESHAW became involved in a scandal for losing nearly $1 billion of investor funds from banks in a derivatives trade. (1997, Bank of America)
4. I am informed and believe that most of the 23 people who were at DESHAW before me are now multi-billionaires.

multidimensional linear regression
SVD (singular value decomposition)
simulated annealing
gradient ascent
gradient descent
Windows GUI
Windows DDK
Stratus
Pascal
Borland C++
C++
C
Unix
Solaris
Sun Microsystems
NASDAQ
NYSE
AMEX
VOS
ISDN
T1
leased lines