The Last Look…

It has to be one of the longest-standing traditions of the FX market – customer asks for a price; dealer makes a price; customer complains the dealer is “reading” them (but deals); dealer complains customer is “everywhere” and dealing away. Let’s roll that into the AI era.

AI will, naturally, learn from its experience, and part of that learning will be along the lines of “customer A interrogates the stream daily at 2pm local and has been a seller 100% of the time they have traded”. What is it likely to do with this information? It probably starts with skewing that customer if possible, depending upon the venue, as well as, potentially, market conditions. Will it have a five-minute view on the market? Probably not, it is more likely to be measured in milliseconds, which is clever, but not, perhaps, optimal in this case.

Obviously, the customer can put the dealer in competition to “keep them honest”, but what happens if all the dealers are using a similar form of AI with a similar level of data? They will all skew, and then it becomes a question, from a best execution perspective, of hitting the least-worst price.

Is it skewing though, or is it ‘reading’? At face value, there is little difference, but if you ask around those who follow codes of conduct and the like (not just in FX), they will tell you ‘reading’ a client puts you into potentially tricky territory. The argument is simple, a skew is something available to all clients on a certain stream, a ‘read’ is something available to one, or perhaps a small bucket of like-minded clients. The former is fair and equitable, the putting a systematic customer at a disadvantage – so goes the argument.

Personally, I have no problem with dealers reading a customer, but then I was a dealer so you’d expect nothing else. I also do not hold with the degree of skew being a factor – you are either allowed to read a client or not, but the latter makes the dealer an Aunt Sally…wait for the smelly stuff to hit the fan and then line them up.

Dealers have always used information in their pricing, but while things have changed radically over the past 10-15 years as CRM tools and data have improved, the advent of AI in pricing could change things even more. Now, allclients’ trading history will be available instantly, there is no need to search or wait for a system to throw out some data, the pricing engine, thanks to AI, will simply skew – or read – the price to an account at a certain time.

Anonymous venues won’t have this problem, of course, but even here, the customer, if they have a pattern, could be spotted, but more than that, on most of these venues top-of-book is probably in one million, when they want to do 10 (or more). No need to skew, just wait for them to hit the price and you know which way the next nine (or more) is going.

If this all sounds cynical, it’s because it is. It is also, to follow up on a previous column, why I welcome the Global FX Committee establishing a working group to look at future challenges, for this will surely be one – and that is because of one factor that seems irreversible in our markets. The number of dealers who are willing to engage, in a public venue, with clients, in what a certain generation would call “good size” (hint: it has more than six noughts kids), has diminished greatly. Back in the day when there were 30 or 40 banks willing to engage, this would not be a problem, most clients could establish multiple relationships and occasionally do what we termed some “back-door trading” to keep their positions away from the same small sub-set of eyes.

There is, of course, a difference between skewing to the wider market because you see the market going that way, and doing it for one client at a specific – and regular – moment in time

Now? Well, it’s very different, as the FX Joint Standing Committee surveys from London show. In the most recent survey, more than 81% of spot business daily involved just six names – in the US survey, just four names handled just short of 72% of daily volume. Historically, this number has been around 75% in London for context and in New York was around that number for five dealers.

So how do we inhibit the ability of AI to read clients? Do we even want to do that? The dealer makes a two-way price, surely they have the right to push it any way they like? As we used to say to our sales desks – they can always hit the other side. The fact is though, that many of them can’t, hence why people are starting to ask if and how we put boundaries on AI-generated pricing.

Obviously, the AI can be given rules that it is not allowed to skew at all, but that inhibits the institution and the technology – and, if the market is moving, who stands in front of a runaway steamroller? There is, of course, a difference between skewing to the wider market because you see the market going that way, and doing it for one client at a specific – and regular – moment in time. I am not sure there should be any intervention unless it goes thoroughly rogue and does things like refuse to quote because it knows the client is asking its AI mates at the same time.

The industry is well aware of the risks of AI going rogue, and is building its infrastructure with these controls in mind, but that is at a macro level. At micro level, are they going to be checking every skew?

A second factor being considered in the industry is does this change customer behaviour? They can, as noted, spread their trading around, but if everyone is using AI, they will get read anyway. If a customer regularly hits multiple dealers, will the AI slow down its quoting as I suggested, or will it skew even more aggressively? Dealers have always been more careful with certain clients (and we can all write down four or five names that have stood the test of time!), it is unlikely to be any different with AI running the show.

Overall, this all seems to point to a world where dealers are put in competition in an anonymous environment, but will they be willing to play there? More pertinently perhaps, will the AI? I regularly recount the story of talking to some astro-physicists at Harvard University about using its super computer to genetically-breed FX trading programmes (this was more than 20 years ago for those that think AI is new). It started out with the best intentions of seeing where the intelligence took it, and within four or five cycles the super computer worked out that some LPs updated slower than others and therefore it became a latency arbitrage programme.

My point is, unlike humans, AI is metronomic and logical – it will see a solution to a problem and act accordingly – in the case at Harvard it was from a user point of view, in this column we are looking at it from a provider point of view, but the tech will process information the same way. The industry is well aware of the risks of AI going rogue, and is building its infrastructure with these controls in mind, but that is at a macro level. At micro level, are they going to be checking every skew? I doubt it.

So clients may have to get used to being read thanks to the limited LP pool available for their type of trading and the fact that the AI seemingly knows everything. They can throw in the odd “confusion” trade perhaps, but how many firms are actually allowed to do that on the buy side, where best execution is, apparently (yes, I am looking at you 4pm Fix) everything?

Ultimately there is good news and bad for clients. The good is that if everyone reads them, their TCA will still show they hit the best price. The bad is, that when they are being read, it’s very hard to argue with an inanimate object, no matter how intelligent it is…although I know a few people who would – myself included.

colin@thefullfx.com