Peek Ahead: DIY Disaster
How one fund house's attempt to build its own algos proved a rocky road to reality.
AT editorial staff spend quite a lot of time listening to the sellside bemoaning the fact that they have to invest an absolute fortune to create and maintain algorithmic trading solutions. One thing that upsets them even more is that some on the buyside are cynically unimpressed. They assume that either (a) the sellside is massively overstating the costs of the exercise, or (b) that they could do better themselves anyway, or (c) that using any sellside algo means you get front run by their prop desk, or (d) all three.
One long-only fund house of AT's acquaintance actually followed through on (b) above and chose to build its own suite of order execution algorithms. (This story has a happy(ish) ending, so the severely disgruntled on the sellside may wish to stop reading now, but the remainder may derive some grim satisfaction from it.)
The fund management firm in question did not have the necessary quantitative skills in house to build algorithms. Recruitment did not go smoothly. The first quant got a better offer and departed after three months. Her replacement (whilst undoubtedly a brilliant physicist) rather lacked the real world touch - as revealed by his tragic assumption that data supplied by vendors/exchanges is always clean and complete. As a result, the catastrophic performance of his first algorithm so highlighted this detachment from reality that it also triggered his detachment from the firm.
The third quant recruited saw the project through, but as a condition of taking the post insisted on a completely different (and far more expensive) development platform from that (briefly) used by his predecessors…
If recruitment proved problematic, technology proved worse, with the major headache being the OMS. This archaic monstrosity had been built in-house by the legendary figure of Reg (the former head of IT) in the mid 90s using "heritage technology". (While Reg always claimed that he had written it using COBOL, his successor's considered opinion was that he'd used Sanskrit.)
Tragically, Reg had nursed his loathsome creation through the introduction of the Euro and Y2K before retiring and almost immediately expiring. To describe what had happened upon his departure as inadequate knowledge transfer doesn't come close - the "documentation" he bequeathed was more a selection of random jottings than a manual.
The obvious conclusion was to buy a new OMS and start from scratch. Unfortunately it was instead decided to incorporate the new algorithms into Reg's appalling legacy - with predictable results. When the algorithms were turned on, the OMS would occasionally run for as long as two minutes before crashing under the message load. Its other party trick was to adjust its order status randomly, so a completed sell order slice would sometimes be added to (rather than deducted from) the outstanding position.
Finally, sanity of a sort prevailed and a new OMS was acquired. Unfortunately, the traders decided to blame the inconvenience of having to learn a new piece of software on the quant. They expressed this dissatisfaction by insisting that a huge and unnecessary selection of additional features be incorporated into the new algorithms. The resulting trader/quant war was unwillingly refereed by the new CIO, who as a new hire had been the obvious victim to lumber with the project's management.
Finally, a year late and more than three times over budget, the algorithms have been successfully integrated into the new OMS. However, the traders are already clamouring for new post RegNMS liquidity aggregating algorithms. The CIO thinks they'll go for buy not build…