Trade affirmations and the coordination problem in capital markets

What makes the trade affirmation process a difficult coordination problem?

By 7 AM, the middle office mailbox of a global macro hedge fund has taken in the overnight events: recaps from dealers in Tokyo and Singapore, chasers from the prime brokers whose SDPs highlighted unmatched trades, and disputes flagged from affirmation platforms that kept running while London slept.

Naturally, most trades, particularly the vanilla ones, have high STP rates and have had their post-trade processing automated for years. These trades remain largely untouched by any human operator and flow through quietly.

What reaches the operator is the remainder: the trades that remained unmatched, had disputed fields, or could not be cleared for some other arcane reason. Dozens or hundreds of items in flight at any one time, no two alike, making up the bulk of the operator’s workday. The queue never holds still: items resolve, new ones arrive, and context switching is the name of the game.

For instance, a CLP interest rate swap has been flagged as disputed on the clearing house field. In order to determine the correct resolution, the operator has to search through thousands of email confirmations, remember that this particular book may have other conventions and confirm it by reviewing historical trades for that PM. This requires both an understanding of the products and systems used by the firm, but also some tacit knowledge and conventions picked up by working with her colleagues and having seen thousands of breaks.

Moreover, context switching happens across teams. At every handover, the context around each open item is rebuilt from scratch, out of forwarded email chains and screenshots: “what is the issue?”, “what has been tried?”, “where is the trade confirmation?”, “has anyone already reached out to the counterparty?” None of it is particular to trade affirmations; the same context reconstruction runs the length of the post-trade lifecycle.

Why has nobody automated it?

There have only ever been two potential solutions, and both have been tried at length: codifying an ever-increasing set of rules or overcoming the cold-start problem of widespread industry standardisation towards canonical trade models.

The first route is more rules: enumerate the exceptions, encode them, and send the remainder to a person. It works only for simple cases that can be easily expressed as deterministic rules. The main difficulty is two-fold. First, the rules need to be able to handle fuzziness in both the fields and thresholds, as well as in the fact that certain rules are difficult to express as a hard rule and require some judgement. Second, each new venue, product and fund convention adds another failure mode. None of them are ever retired. A rulebook can only describe the failures somebody has already seen and written down, and the team is dealing with new ones that arrived this morning.

The second route is widescale industry standardisation. A common schema that both counterparties conform to would simplify the problem. But making that schema a precondition requires a critical mass of participants to move at the same time, and coordination on that scale has rarely been achieved in this market. Even where a platform exists, the process leaks around it. A trade affirmed manually outside the platform cannot be novated inside it, because the platform holds no record of the item being transferred.

Enter Arch

The latest advances in AI now mean that the judgment calls in these gaps - whether a recap describes the trade the firm booked or whether a difference is economic - can now be made by an agent by collating and reasoning over information scattered across systems.

Importantly, such a system needs to have a few key features. Specifically, the ability to:

• Construct context by accessing all of the relevant systems. Search through, collate and analyse relevant artefacts and evidence associated with a particular trade by accessing trade databases, affirmation platforms, mailboxes and other document stores.
• Escalate to an operator when insufficient data is available or conflicting evidence exists. Apply clear guardrails that define when the available evidence is sufficient to propose an action, and when further analysis or input is required.
• Continuously learn from new breaks, feedback and inferred resolutions. Conventions that have never had a home outside the operators’ heads get written down as a by-product of the work, thereby making the problem tractable.
• Execute actions with human approval or autonomously, as configured by the team. Taking the action proposed by the agent - a prepared email to the counterparty, a short message to the PM, a payload to amend a trade in the OMS - should happen with a click. As the trust in the system is earned, configurable autonomy ensures that low risk actions can be taken autonomously.
• Fully audit the analysis, recommendations and actions. The system needs to provide clear reasoning at every step, make all the evidence (trade data, communications, etc.) readily available, and enable an operator to see every single tool call (e.g. SQL query, Outlook message search) that has contributed to a particular action.

Arch is Yantra’s capital markets operations platform; T0 affirmations is one of the workflows it is successfully handling today. It runs as a set of agents that listen to the trade flow continuously. The affirmation platforms show every unmatched and broken trade in real time, and that picture moves as submissions, amendments, terminations, novations and allocations land against it. Additional information keeps arriving after the initial break - a second recap, a correction from the broker, an allocation that turns up late - and any of it can change what the right resolution is. The agent resolving the break keeps listening and proposing updated actions.

The boundary between proposal and action is configured by the team rather than assumed by the software, and it is meant to evolve. Once a category of break has been resolved the same way often enough, the team can let the agent execute the actions autonomously and review it ex-post.

None of this requires the market to agree a standard; the work was waiting for something that could sit at the edge of the existing systems, with relevant organisational context, and act with the operator still in control. These agents are the first step towards circumventing the coordination problem that has held back innovation in market infrastructure.