Your RFQ responses already rank your dealers. No one's looked.

Execution desks make dealer selection decisions dozens of times a day on the basis of relationship and instinct. There is no systematic record of counterparty pricing behaviour because, until recently, there was no way to build one. The data exists. It just arrives in a form that disappears the moment the conversation ends.

The data inside every RFQ response

Every RFQ your desk sends generates a response set. Counterparty, instrument, level, speed, whether they were hit or not. Across a week of activity, that is a detailed picture of who prices competitively on which names, who responds inside a minute and who takes ten, which dealers quote aggressively on a name and don't get hit and are likely still axed in that direction.

That intelligence is quantitative. It arrives as conversation, embedded in chat threads, and the archive grows while the analytical value inside it goes untouched.

The same conversations contain axe flow: dealers and counterparties indicating interest or inventory in specific names. An execution desk that captures axe positions systematically has a real-time picture of who is motivated to trade. Most desks get a fraction of it, from whoever happened to be watching the right chat room at the right time.

Why has the data stayed inaccessible

Two things have kept it out of reach.

The first is API access. OTC chat platforms kept conversation data locked inside closed systems until recently. No extraction, no integration, no way to connect to what was happening inside the platform. Vendors have opened APIs, and it is now possible to connect to those conversations programmatically.

The second problem is harder. A conversation between two traders on a rates desk is not legible to a general AI model. The language is domain-specific: abbreviations, shorthand, conventions that vary by asset class and by desk. Those models are trained on publicly available text, and OTC trading conversations have been private by nature. A model that has never encountered the way a credit desk talks about a swaption run will not reliably extract the fields that make the data useful.

Why fine-tuning doesn't hold

The standard approach to this problem has been fine-tuning: label a corpus of trading conversations manually, train the model to recognise the patterns, and deploy. It works for a period and then degrades as the desk evolves — a new trader joins with different conventions, a product line shifts — and the retraining cycle starts again.

There is a more fundamental problem. The people doing the labelling are not the people who had the conversations. A data scientist working through a credit derivatives transcript does not know the prior context between those counterparties. The label can be wrong before training even starts.

The model that avoids this learns from the trader who did have the conversation, in the workflow they are already using. ipushpull captures the conversation, classifies the messages, extracts the structured fields and surfaces a draft record for the trader to review. Where the trader corrects the output, that correction feeds back into the model immediately. Accuracy improves over time. The desk's evolving language and products are the training mechanism, not a constraint on it.

How it runs in practice

Your data remains yours. ipushpull is the tool that extracts it for you, but everything is completely private and never shared beyond your firm. All the training is kept inside your platform, only to be used by your own desks, ensuring no leakage can occur and we're not using your traders to train your competitors' advantage.

Every inference is logged: what input was given, what fields were extracted, what changed in the review. Permissions are field-level: the right people see the right outputs without the underlying conversation being exposed.

Quote Hub runs on this basis at a number of the world's top ten hedge funds. The counterparty analytics it produces come from each firm's own activity, scoped entirely to that firm. No external benchmarks. No data is shared across clients.

Dealer selection decisions that were based on habit can be grounded in a month's worth of actual response data. The question of who to include in the next RFQ has an answer that does not depend on who happened to remember which conversation.

The record was always there

Every RFQ your desk has sent in the last year generated data: who responded, at what level, for which product. Dealer nine has been the most competitive on this name all week. The record of that has been sitting in the archive since the conversation happened. Quote Hub makes it queryable. Learn more about Quote Hub.