Profile : Santiago Braje : Katana

Bringing AI from the sell side to the buy side

The Science of Investment spoke with Santiago Braje, CEO of Katana, a trade idea platform for bond markets, about the capacity for investment managers to employ AI tools.

Katana was initially a sell-side pricing tool, how did it transition to a buy-side tool?

The tool that we launched with our company is a spin-off of the first project. We are not using anything in terms of the technology, or the data that we were using for the internal project, we started from scratch.

We believe there is a huge opportunity now to use AI/machine learning and data science. We specialise in decision making for investment management and in fixed income, particularly. We understand the decision-making problems in this space as complex choices. There is room for automated trading algorithms and systematic strategies functioning on their own, but we think there is also room, and perhaps the greatest opportunity in using that power, to make the decision process more robust which ultimately continues to be controlled by a human.

What is the opportunity that you see?

More money is being spent on the execution side, which is the last stage of that workflow. That has become very competitive, and in some parts of the market you have big players like Bloomberg, Tradeweb or MarketAxess.

There are also firms that operate ahead of the execution process like RSRCHXchange, aggregators for analysis to make that consumption more efficient. What we haven’t seen much is in the areas of discovery. The PM still needs to be deciding which bonds to add or to sell from the portfolio. That process involves several steps, and machine learning embedded in the process can play a big role in making that process more efficient and more effective.

We approach that problem from the perspective of rationality. When managing money you definitely want to be rational. We know how rationally human beings should behave, we know they don’t behave like that.

The basic idea is that you will look at all of your available choices and you will compare them in an objective way, to ascertain which is the best one.

The other aspect is that when you have money in your portfolio you are looking for value; value is always relative. Every time a portfolio manager chooses to buy a bond because of its yield, it is set against the yield of that bond compared with all the other bonds that they could buy, or against another benchmark.

How are you achieving that?

In one solution, we look at relative value, and also at the choices you can make. The system can do that by looking at all the possible bets that you could make in a given universe. So, if you have 2,500 bonds you will have 3 million possible relative value trades. The PM’s time to analyse a potential opportunity is limited, so a rational process should start by looking at all of them to determine which are the relevant ones.

So instead of comparing one bond against the index of 2,500 bonds, you compare each one to all the other bonds and find the ones that actually have similar behaviour, or that are closest to the one that you are analysing and then whatever you observe a deviation it will be more meaningful, because there is a closer correlation effectively between the behaviour of those two bonds.

It gives you an indication of how closely related two bonds may be. If we do that for the whole universe, so you have 3 million combinations. Then essentially you can run the algorithm for every pair of bonds and see which ones are more relevant for making estimates about the future behaviour of that particular bond.

Then we look at where there are anomalies, and through this process we go from 3 million possibilities to 100. So, it’s a big filtering exercise. If you are running an investment process and you start from there, you immediately go to things that have a high likelihood of being mispriced or happen to be dislocated. And you do it in a framework that is agnostic, and complete in a sense that it’s looking at all the possibilities.

Where is this fitting into the workflow between traders and portfolio managers?

We provide a product that does not need any integration, we have our own data supplier that is IHS Markit, and any asset manager can start using it straightaway. Delivery of the product is through web browser, and all you need is that we set you up with an account.

Different firms use it in different ways. For example, one asset manager here in London has their own in-house relative value models. Their own analysts follow the company, and they rank the company by relative value. The PMs and traders all see that. Katana is a perfect complement for that, regardless of who is using it. It’s easy for multiple people to use it and makes it easier for people to communicate and share ideas between PMs and execution traders, because they are all seeing the same thing.

In the case of this firm, they regularly have that two-way communication flow, where maybe the trader finds out something interesting and brings it to the PM, or the other way around where the PM will initiate the idea. We have the same kind of thinking. There is no reason why it shouldn’t be like that, the PM or the trader need the tools to communicate that effectively. Ideally what we provide helps the team to collaborate more effectively around the idea, so it’s more informed, it’s less about an opinion and it’s more quantified.

Can your platform spot opportunities and sends alerts to the PM/trader, or does it require them to pull the information when they need it?

Both, based on how it is customised. For example, today in an EM universe there are 2.5 thousand bonds in the JP Morgan index, and we are producing 100 ideas. The machine can spot where there is a dislocation. When you see the change in the relative value of the 2 bonds, it’s pretty obvious that the dislocation is there, but to find it you will need to run through three million possible comparisons. That’s what the algorithm is doing.

There is this agnostic generation, and you will have different kinds of trades. You have different countries and some that are let’s say there is a switch between two countries for example Angola and Ghana. Others will be in the curve. If you are a PM you may have some focus areas for example, so maybe 100 ideas are too much, the idea is that you quickly find what is more interesting to you.

How many categories do you have?

So we have five filters, senior/sub, country, industry, ratings and maturity. Development has been user-generated, so each new feature included was generated from users’ feedback and interaction with the tool. For example, filters allow the PM to go as fast as possible to the area he or she focuses on, or can search for specific issuers or bonds.

What are the communication features?

There are things like bookmarking, so if I find something interesting a PM can put it in their bookmark so it will show in the trader’s bookmark also. Then they can communicate directly, to say ‘check this one’ or share perspectives. It makes it really easy to interact.

The other thing that we are doing, in terms of pushing and pulling, is search functionality. Say you have a bond that you want to sell, we may not be picking up any relevant ideas as there may be nothing interesting happening. You may want to sell for some other reasons; you may have a portfolio constraint, mandate or risk concern. We can manage that, and we do best relative value ideas for that specific bond. From a statistical perspective they will not be as strong as if they were agnostically selected, but they will still have better expected performance than anything else.

That is already available. We just launched in November and we are adding more features and refining the product now. This is an area where we are making it more customised, and easier to use.

Is that user led?

We see a lot of users do that. They start searching for all the bonds they own, for example. We can do that in one go. Then you start to see some more interesting things at a meta data level. For example, when you compare one thing against another, rather than just looking in isolation. If you are doing this exercise of looking at all the pairs and then looking at which of those end up being good, then you find that some bonds end up being a good idea against many others. That is telling you, with a lot of accuracy that a bond has moved in an unexpected way versus many others.

This will change a lot on a daily basis, so today we might a lot of change in Sri Lanka for example, last year we saw the reaction to the Chile riots. So it’s interesting, we are using this lens, pulling things in pairs, you open all kinds of possibilities for getting to grips with what’s happening in the market generally.

How would you characterise what is being studied?

It is really looking at this from a perspective of what has changed, and which of those changes are relevant, or not. Everything is changing all the time, so it’s not a trivial problem. You are only coming from that aspect of it, so you don’t know if there are structural differences. However, you would expect that structural differences are already digested by the market, and this should be embedded in prices after some time. However, if there is a change that is relevant, we will pick it up.

Could this be used to build an asset management business?

We get asked a lot of the time why we are selling this and not running our own asset management firm. It’s because we understand that you can have a tool that is fantastic at spotting the relevant changes, but that doesn’t mean that you are able to discern which changes are actually worth pursuing, and which are the ones that are not. We know the PM has a role in providing subject matter expertise to make the call as to why this move took place. Maybe a bond came up with a tender; that would change the price, but that doesn’t mean that there is an opportunity there. We see that there are things we can do to make those decisions easier, and eventually you will get closer and closer to cover more sort of aspects of the decision with machines. But making the leap that, because we can identify anomalies you will immediately have a successful, consistently successful asset management business that is too far.

What concerns do your clients have around using AI?

Trust is shifting from being a bilateral thing to more that you have to trust your instruments. Pilots are in command of their aircraft for about 2 per cent of the time on average. They trust the computer for the other 98 per cent. We see that shift. Clients are seeing it, trying it out by going out and actually executing trades, in smaller sizes. Then over time we are seeing people doing bigger sizes. There’s a transition, an increasing level of trust where at some point it clicks and instead of pushing back they start to say, ‘Give me more.’

Do they have any concerns about understanding the logic or retrospectively looking at why decisions were made?

We are not making decisions for them, that reduces the issue. A lot. We also interact a lot with the users to get that feedback to learn, where we are sort of flagging things that we shouldn’t be flagging, or where there are things that don’t work very well. It is a statistical problem, you want to balance the false positive and false negatives.

Is data quality a concern?

The market is becoming more transparent and pricing is becoming better. Is there more to do? Definitely. There is a lot of discussion at the moment in terms of creating a consolidated pricing tape for Europe. I think the market will welcome that. There is also value in having a central price reference for the market. But price accuracy cannot be detached from the fundamental liquidity of instruments. Fixed income is a very broad universe, our data suppliers there is over 100,000 securities of pricing. To be meaningful, prices need to reflect transactions. With so many instruments out there, portfolios are not turned too often. If you look at a typical fixed income asset manager will turn a portfolio once a year or less. So, there aren’t that many transactions to get prices better. There is so far you can go. That doesn’t mean that things cannot improve.

How do you bridge the gap where data is not available?

We use reference data from IHS Markit, and for the bonds that are more normally traded, everything that is in an index, you will tend to have some pricing data. Dealers that will be sending runs, and contributing prices to pricing aggregators like MarketAxess or Bloomberg, etc.

So there is some data there, and that is for what we found at least in the markets we have been looking at. That is good enough, even if you are doing an end of day which we are using at the moment, to give you a context of where things are. It’s always a matter of size, so if you keep increasing the size then that liquidity and transparency disappears. At that point you go past price discovery and engage in liquidity discovery. ●

Biography: Santiago Braje began his career in corporate finance at Societe Generale and Citigroup in Argentina. He joined ING Financial Markets London in 2005 and became a managing director in 2009. At ING, he led emerging markets and structured credit trading before taking responsibility for all credit trading globally. From 2017 he was also the business sponsor of Katana, raising capital from ING Innovation Fund over several rounds, and in November 2019 left ING to found Katana Labs Ltd. He holds a BA in Economics from University of Buenos Aires, an MSc in Economics from Torcuato Di Tella University and an MSc in Operation Research from the London School of Economics and Political Science.


Leave a Reply

Your email address will not be published. Required fields are marked *