Everyone knows that SecuritiesDatabase is Goldman's special strength. That the system, otherwise known as SecDB, which allows Goldman to price trades and assess risk for 2.8 million+ trading positions daily, is a huge source of competitive advantage and something other banks are trying to emulate. But what if SecDB's strength - that it's years ahead of rival systems - is also its weakness? Some in the industry are quietly suggesting this might be so.
We didn't speak to anyone at Goldman Sachs in an official context for this article. We did speak to current and ex-technology staff on an anonymous basis. Some suggested that the SecDB edifice is inherently rickety.
"You have to understand that SecDB is an amazing platform with tons of genius ideas, but that it's 26 or 27 years old," says one former Goldman insider. "There were a lot of technology choices that were great for the business in the 1990s and 2000s, but things are different now. Back then, for example, you might have had a couple of hundreds of thousands of trades a day, but now - in electronic FX trading - you might have a million trades a day. How do you scale a system that was built to work with 100,000s of thousand of trades to millions? This is the problem for GS."
Understanding SecDB requires an understanding of Goldman's history. Along with Goldman CEO Lloyd Blankfein, CFO Harvey Schwartz and securities heads Pablo Salame and Isabelle Ealet, SecDB originated with J Aron, the FX and commodities broker acquired by Goldman in 1981. J Aron wasn't fully assimilated by Goldman until the early 2000s, allowing it continue developing its own technology. "J Aron started using SecDB around 1992," says the ex-GS insider. "It was maybe 2006 or 2007 before the whole firm was running it- every time there was a crisis like Asia in 1998 or the dotcom crash in 2002, J Aron's SecDB system was extended a little further."
From the outset, SecDB was a collaborative effort between J Aron's "strategies" team (the etymology of Goldman's contemporary 'strats' group) and its technologists. Strategies was run by Armen Avanessians, now head of Goldman Sachs Asset Management’s quantitative trading business. The technology side was managed by Mike Dubno, Glenn Gribble and Kevin Lundeen. All were very talented technologists, but this was 1991: many of today's programming languages simply didn't exist.
In the absence of Python, Goldman's tech people invented their own language: Slang.
"It was an ingenious idea," says another ex-Goldman insider. "There was this realization that a language like C++ is good for execution speed - when it runs, it runs really quickly - but that it's a complicated, messy language which takes a long time to develop. The technologists figured they needed a language that was easier to use - an interpretive language that could run across Windows, Solaris and Unix systems. So they came up with Slang."
Goldman still uses Slang to code SecDB this day. According to various estimates, the system is built on anything from 15m to 40m lines of Slang code. Slang has been good to Goldman Sachs, but now there's something better: Python.
"The other banks which are developing systems like SecDB now are using Python," says the ex-GS technologist. "With Python, you can leverage what everyone else is doing and take advantage of things like TensorFlow [Google's open source software library for machine learning.] Goldman can't do that - because of Slang, it's stuck doing everything in-house. With Slang, you can't propagate ideas from elsewhere. Ultimately it will reduce Goldman's pace of innovation."
The senior technologists suggest Goldman's reliance on Slang hasn't been much of an issue so far because opensource code libraries haven't had much of relevance to banks' trading systems. As artificial intelligence comes to trading this is changing: in future, Goldman will derive bigger advantages from leveraging AI-related Python code developed elsewhere.
So, why not move away Slang now? It's not that easy: those 15m+ lines of code can't be replaced overnight: it's an undertaking that could take a decade or more.
Other Goldman insiders say the firm's aware of the problems and has been addressing them for years. One ex-Goldman partner says the issue of scalability was solved 17 years ago when SecDB was rolled out in equities: "We *knew* in 2000 that the volume of trades was going to increase. It was the beginning of the rise of electronic trading. So we solved the scalability problem then." Others say Goldman already has "JSI" or Java Slang Integration - a method of accessing SecDB using Java code, and has developed Python wrappers.
The shift to Java in particular is thought to be related to the comparative ease of hiring Java programmers. This too is another problem with Slang: unless you've grown up at Goldman, you won't be familiar with it. And if you work in Slang at Goldman, you'll fall behind on all the innovations in languages like Python. "It makes it harder to hire laterally," admits a Goldman technology insider. This is one good reason why Goldman needs to hire so many STEM students: it has to train Slang programmers from bottom up.
For all the issues with Slang and SecDB, Goldman technologists past and present emphasize that the firm remains the best place to work as an engineer. "The amazing thing about Goldman and SecDB is that it's a platform which allows developers access to integrated data and access to parallel computers and tools for collaborating with colleagues on writing code and deploying code to products in a controlled way," says one technologist at the firm. "The whole software development life cycle is in one place. You have a team of developers at GS who own the stack right up to traders' desks, who understand the whole business. By comparison, anywhere else is incredibly bureaucratic and inefficient."
Nor is learning Slang held to be that onerous. "A good programmer will pick up Slang easily. It doesn't matter that you're not using Python," says the engineer. "Yes, SecDB has its limitations, but it's held up for 25 years so it must be good."
As Goldman rolls out SecDB to its clients in the form of Marquee, the firm needs to hope this sentiment is correct. After all, banks like J.P. Morgan - whose Athena system is coded in Python and is also being extended to clients - are snapping at Goldman's heels. As we reported yesterday, J.P. Morgan has also hired one of Microsoft's top artificial intelligence professionals to develop a new global machine learning division. If Goldman's ex-technologists are right, it's in the interface between SecDB and artificial intelligence that Goldman could find itself most exposed. And this is the weakness that J.P. Morgan seems to be going after.