Galen Stops takes a deep-dive look at how Goldman Sachs is adding self-learning elements to its FX algos, and where this technology fits into the shifting analytics landscape across FICC markets.
Goldman Sachs has for some time now offered a variety of algorithms to clients for trading FX, ranging from the more aggressive ones that are designed to quickly sweep the available liquidity to the more passive ones that focus on how and where to place interest in the market.
Sitting between these two extremes is the dynamic hybrid algo which is, according to Ralf Donner, head of FICC execution solutions at Goldman Sachs, “partly passive and then cautiously aggressive”. The challenge with such a product is clearly that, between these approaches, the algo needs to be able to determine how it should be behaving at any given point in time based on a variety of market conditions and execution goals.
As Donner explains: “If you look at something like a liquidity gauge for FX markets intra-day, then you observe that spreads undergo widening and contraction, sometimes for hour–long or multi-hour periods. For example, you might have more structural issues like spreads that widen out significantly during Asian time zones compared to European ones. Meanwhile, you have effects that are taking place on a much shorter time scale, such as around fixings or a news event, where the order book behaviour suddenly changes in just a matter of minutes. Now if you try to take all of this into account and build an algo that solves for all of these problems simultaneously in a pre-determined, pre-calibrated way, it won’t work. We just don’t think that it will be able to solve for all of those situations. So the alternative is to have an algo that is self-adjusting or self-learning.”
This is exactly what the Goldman Sachs team has developed.
It’s worth noting that staff there are quick to shy away from terms such as “artificial intelligence” (AI) when describing what they’ve developed, partly because it has been reduced to a buzzword in some quarters and partly because it doesn’t particularly help illuminate how the technology works. After all, as Donner quips, this algo isn’t going to pass the Turing test.
Measuring the benefits
Instead, they prefer to just describe it as an algo that looks at its own performance and the current market, then looks at how it is doing versus its target and adjusts its behaviour accordingly.
“Historically, we would have an aggressive ratio that we wanted to target in the market and we would periodically revisit it as spreads change and market structure changes. The benefit of this technology is that you’re able to target a real-time aggressiveness ratio and as your algo exceeds or is below that ratio, it has the ability to self-correct so that it’s more in line with its target behaviour,” says Swayam Thacker, head of FICC execution solutions Americas at Goldman Sachs.
But how to measure the benefits of such a self-learning algo?
Well, one of the key considerations for the algo, say Donner and Thacker, is how often it should be crossing a market top-of-book price. For example, for a fairly passive approach the target could be set fairly low, say at 10% of the time.
What the Goldman Sachs algo team found was that an algo using a static model calibrated on a periodic basis using a fixed parameter, such as distance to mid, to aggress liquidity produced a lot of variance from the pre-determined target. So instead of being aggressive 10% of the time, the algo might have instead ended up being aggressive for 5% or 20% of the time as it executed instead.
By contrast, they found that the new self-learning model led to a tightening of this variance, in addition to greater accuracy for the target agress percentage intended for the algo. In other words, the algo was more likely to be aggressive 10% of the time, and even if it misses this target, it is likely to do so by only a small percentage.
The next step forward
The other consideration is what adding this self-learning capability does in terms of execution outcome. To figure this out, the Goldman Sachs team looked at what using this algo did for the implementation shortfall on execution, what it did for performance versus a risk transfer price, etc, and they found that the algo performed well compared to the benchmarks they used.
An additional benefit of the self-learning algo, says Thacker, is that the average duration that the algo needs to execute an order shrank. This is because with the traditional static model if markets suddenly widened out, the algo could struggle because it didn’t understand that the real-time market was actually much wider than it had been trained with historically, whereas the self-adapting algo can adapt in real-time to adjust to such market conditions and aggress appropriately, which in turn reduces the overall length of time of the algo.
“In a broader sense, it’s been very popular for banks to produce pre-trade tools for clients to look at analytics and be able to assess from market conditions what algo might have been appropriate. The next step on from that is to embed some of the findings from the pre-trade tools into the algo logic itself, and we see this as a step towards doing this,’ says Donner.
One of the obvious questions surrounding any type of algorithm that has logic built into it in order to take decisions in-flight based on market conditions is how it will fare with extreme events such as a flash crash or a major market disruption, à la SNB.
One of the challenges for self-driving cars is that they can be fed all the data in the world, but if they encounter a real world situation for which there is no pre-existing data to guide them, they can struggle. On the surface, the situation here seems analogous because, if the algos are designed using existing data, how can they be expected to cope with some kind of market dislocation for which there is no previous data?
“That was a crucial question that we had to address when designing this model – how does it perform during NFP, FOMC time or when there’s an event like the yen flash crash? How does the algo react and how does it adjust itself?” says Thacker.
He continues: “What we have done is put protections in place to handle such extreme events in an appropriate manner. For starters, there are bounds in place as to the level of aggressiveness ratio that it can target in the market, and this is dictated by general market conditions and the spreads that we see.”
What this means in practice is that the algo looks at what the primary markets are doing and imposes caps on how aggressive it can be based on that. This kind of limitation is mainly relevant for instances when the FX market moves wider than usual, but still remains within what would broadly be viewed as normal conditions.
“Then there are other caps that prevent this algo from doing something damaging in the event that the primary market really blows out or market conditions become really unusual,” adds Donner. “And those are hard limits in actual numerical terms on how far from mid the algo may aggress.”
While the Goldman Sachs FX algo team see limited potential applications for AI in that market, their remit was expanded earlier this year as the bank created a new “FICC Execution Solutions” team, which also covers futures, and looking at these other markets they see greater opportunities for this technology to be applied.
“While we don’t yet see any strict application of AI to the FX world, and in general we would question whether or not there is a genuine application at all as of yet, we do think that there are potential applications in these other asset classes. When you have an instance where an algo that you’ve built needs to be able to execute thousands of different instruments and there’s no way of working out for each instrument what its natural characteristics are, then AI beckons,” says Donner.
As Thacker points out, in FX the G10 complex comprises 45 currency pairs, whereas in the futures market there are thousands of different instruments, many of which are not even liquid.
“In the futures market you find that some of the hallmarks of some of the contracts can be applied to other futures contracts. But in FX you have the luxury that each currency pair can be carefully studied and therefore using a more supervised approach can actually be beneficial because you can understand that currency pair in much more detail,” he adds.
Looking at the creation of this new FICC execution team more broadly, the logic behind it is twofold.
Firstly, there is a belief that the asset classes are similar enough that different teams across the bank could learn from each other if they were less siloed. Yes, some of the order books might look and feel a little different in some cases, but the firm’s clients are often dealing with similar challenges across each of these asset classes.
Secondly, Goldman sees an opportunity to build transaction cost analysis (TCA) tools that can be applied commonly across FICC markets, whether that is pre-trade analytics, an algo wizard to help clients pick the most effective algo for different circumstances and execution goals, intra-trade tools or post-trade analysis.
A new battleground
Indeed, Donner says that there is still a long potential runway for development of FX TCA tools across these markets, although the landscape is still shifting.
“I think that the post-trade project has been abandoned by banks in favor of third party solutions. But to be clear, when I say “abandoned” I don’t mean that they’ve stopped producing post-trade reports, it’s just that if you’re a confident algo provider, then you welcome third party TCA. So in general, that’s now an outsourced problem,” he says.
Donner continues: “There’s still a lot of interest in pre-trade TCA because it’s difficult to build generic pre-trade tools, they need to be tailored to the algos themselves, and so there is still scope for banks to be doing their own pre-trade analysis. But now the battleground has shifted to providing tools that can offer some useful information to clients intra-trade. This can be done through single–dealer platforms, but we actually launched this on Bloomberg recently through a fully integrated solution with the multi-dealers where the analytics pop–up immediately together with the Bloomberg algo ticket and then the client is able to see their execution. This intra-trade element is very recent and it’s probably going to be a continued area of focus.”
It seems likely that the TCA battlegrounds of today in FX are likely to become the battlegrounds of tomorrow across the rest of the FICC markets.
“Futures is a very electronically traded asset class, but if you look at the other cash FICC markets, such as Treasuries, then algos are very much only in their early stages. The market structure there is different and there are more products on the illiquid side of the spectrum, but electronic order books are becoming increasingly prevalent across all of FICC and as the liquidity pools there become more electronic, what we’re finding is that clients want to take ideas from FX and futures and apply them to these other asset classes,” says Thacker.