Earlier this month, ICE launched ICE Compass, an AI-driven pre-trade analytics platform that ranks dealers on the competitiveness of their pricing and estimates where a fixed income trade is likely to clear before the buy side signals any intention to transact. T. Rowe Price, which provided feedback through development and beta testing, has signed on as anchor client, with a further four or five firms described as in active discussion.
What the platform targets is a long-standing exposure on the buy-side desk. In a market still built on bilateral negotiation rather than a central order book, asking a dealer for a price is also the act of revealing intent, and that disclosure carries a cost. Show an order to the wrong counterparty, or to too many, and the price can move before the desk has committed to anything. Compass applies AI to a firm’s own data to narrow the distance between asking and trading.The premise behind the product is the now-familiar problem of abundance. “The buy side is bombarded with an enormous volume of data, and that volume is only going to increase, not decrease,” says Varun Pawar, Chief Product Officer, Data Services at ICE, in conversation with TradingTech Insight. The electronification of fixed income, the growth of portfolio and basket trading, and the daily flow of bids, offers and indications of interest from dealers have left trading desks with more raw material than they can readily interrogate.
The gap Compass targets is the capacity to turn that data into decisions. “The majority of buy-side firms we’ve spoken to are data-rich but have done little to extract insights from that data, not only on an ongoing basis but also historically,” says Pawar. Some of the largest and most sophisticated desks have built this capability in house, having the engineering resources to do so. Most have not.
What the model estimates
Compass works on a per-instrument basis and in near real time. A desk can put up a list of bonds and see which dealers have quoted on each that day, alongside a score for each counterparty. A dealer’s screen price, however, is the opening position in a negotiation rather than the outcome, and it is that distinction the model is built to capture.
“Based on the data coming in from the various dealers, along with the historical trades the firm has done with those counterparties, we can produce a predictive estimate that is more realistic of where they can expect to trade,” says Pawar. In practice, the model estimates expected slippage dealer by dealer, calibrated to how each counterparty has actually behaved with that specific firm rather than to a market-wide average.
The same dealer may rank differently for two different clients, because the inputs are each firm’s own quote flow and trade history rather than a shared benchmark. “It is important to note that the algorithm is calibrated specifically to the firm. We’re not applying generic numbers across the street,” says Pawar.
A per-client architecture, not a data pool
The data design is where Compass departs from the consortium model that pre-trade analytics in fixed income often implies. Three sources feed the model: a firm’s historical trade data, ICE’s pricing and reference data including its continuous evaluated pricing, and the daily quote and indication-of-interest flow the firm receives from its counterparties.“We are deploying algorithms and ICE’s data into the client’s instance, and all of that information is proprietary to that specific buy-side firm,” says Pawar. ICE says there is no crossover of data between Compass clients, and that none of the underlying buy-side or counterparty data returns to the company. The pooling the platform performs is the aggregation of a single client’s many dealer relationships into one view, not the pooling of information across the firms that use it.
That containment matters for a function whose purpose is to limit disclosure. A desk gains a sharper read on its counterparties without its trading footprint surfacing in a dataset that rivals, or its own dealers, might draw on.
A feedback loop into the sell side
The intelligence is not designed to remain on the buy side alone. The same rankings that inform a desk’s order routing become a record it can take into periodic dealer reviews. “ICE Compass gives buy-side clients the ability to have more meaningful data-driven conversations with their sell-side counterparties,” says Pawar. “This intelligence allows dealers to understand where they rank relative to others and, more importantly, areas they should focus on to stay competitive.”
The mechanism improves with use. A client first supplies historical data so the algorithm can be trained and calibrated, then feeds quotes and trades intraday, with the model retrained incrementally as the day’s activity is incorporated. The post-trade record becomes the input that sharpens the next pre-trade estimate.
Measuring the impact
The benefits ICE describes are framed as the outcomes the platform is built to deliver: a reduction in information leakage through more assertive dealer targeting, and lower transaction costs through better dealer selection or a change in trading strategy. As an early-stage launch with one named anchor client, the platform’s track record across a wider set of desks is still being established, and quantified execution results will accumulate as adoption grows.
Best execution gives the proposition a clear footing in the meantime. Buy-side firms are obliged to obtain the best available outcome for their clients, and a tool that brings additional data and rigour to dealer selection maps directly onto that requirement, offering desks a structured way to evidence the obligation they already carry.
For ICE, the launch is also a marker of direction. A firm long defined by its evaluated pricing and reference data is moving from supplying the data that desks consume to operating the analytics layer in which decisions are made. Compass is one product, but the shift it represents – from data provision to embedded, AI-driven decision support – is what firms will be watching.
Subscribe to our newsletter
