For asset management firms investing in the primary market for corporate bonds, achieving a complete, real-time view of new deals coming to market and changes to deal terms is key to efficient credit analysis and rapidly placing orders for allocations.

However, syndicating banks use multiple and competing channels to disseminate data to asset managers, making it challenging to aggregate and reconcile key information.  While deal platforms like DirectBooks and Ipreo (S&P Global) are important sources, emails and instant messages still play a major role in communicating deal data to asset managers.  With each deal receiving up to 30 updates across the duration of the syndication process, it’s difficult for trading desks to stay on top of the latest updates and keep portfolio managers informed.

Unlike data delivered by platforms, email and chat messages are unstructured – a syndicate bank   can use a limitless variety of data formats, terms, data labels or naming conventions. In part, variability is exacerbated because the messages are generated by individuals, rather than systems on the bank side, and they use their preferred formats and terminology for variables like tenor, callability, coupon and currency.  Historically, it has been difficult, if not impossible, to auto process these data, partly because traditional parsing techniques cannot cope with the variability in deal messages. 

As a result, asset managers have had to use time-consuming, error-prone, manual workarounds to manage off-platform information, until now.  AI enables the creation of more powerful, adaptable data processors and is uniquely suited to interpreting and extracting deal information from the unstructured data contained within emails and chat messages. 

Most AI relies on a large language model (LLM) to understand and extract meaning from text.  Training an LLM for a specific task is complex and LLMs can produce variable results from the same inputs.  However, careful fine-tuning for specific use cases can yield highly accurate results, making a trained LLM suitable for interpreting primary market bond data.

Despite rapid innovation in public LLMs, we believe that a private LLM is best suited for most use cases in institutional capital markets.  With a private model, it is easier to safeguard data, directly train the model on a specific task, calibrate its performance and control costs.

Using AI for email and chat message processing provides a potent tool for achieving a comprehensive view of the new deal market spanning on- and off-platform deals.  Aggregating deal data enables the creation of an integrated, deal-focused workspace that optimizes how asset managers operate in primary markets for corporate bonds.

In addition, the ability to process unstructured deal data can provide a more real-time view of a deal and the market.  For example, a syndicate bank might transmit a change in coupon or other key term via email or chat before updating the deal on a platform.  In that scenario, the AI-driven system could update an asset manager’s deal screen before the update is published on the platform.  Similarly, AI can process grey market data to bring additional intelligence about the demand for specific new deals.  When pricing windows are open for only a few hours, any advantage in making decisions and placing orders for allocations faster should benefit the asset manager.

In financial services, new technologies often hunt for problems to solve. Sometimes, the hype exceeds its practical value, especially in the early days.  AI is different.  We believe AI can improve existing technology and enable automation where it was previously untenable.  Accurately and efficiently processing unstructured data is a good example of how AI solves a long-standing challenge in primary bond markets and how this technology is poised to pay dividends to proactive asset managers.