This is the first technical note we are publishing as a company, and it is appropriate that it covers the core question: what exactly is the Deep MM fair value signal, what data do we use to produce it, and what are the accuracy targets we are working toward. This note is a preview, not a validation report. The model is in active development and the numbers we describe here are targets and early-stage estimates, not established facts. We publish this to be clear about what we are building and to give institutional desks a basis for evaluating whether the approach is one they would want to work with.
Why We Started This
The problem that motivated Deep MM is straightforward to describe but persistent enough that it has not been solved at scale: the majority of corporate bonds trade so infrequently that a fixed-income desk cannot get a reliable current fair value from market data alone. TRACE data covers trades as they happen, but for bonds that do not trade on a given day or week, TRACE is silent. Dealers provide indicative axes, but those axes reflect inventory positioning and dealer appetite, not an unbiased estimate of where a bond would trade if someone wanted to do size today.
A portfolio manager running a large IG book at the end of the month needs marks for hundreds of positions, many of which have not printed in weeks. Their options are to use the last observed trade (which may be far from the current market), to use a vendor's model price (which is often a simple duration-matched interpolation that ignores the bond's specific characteristics), or to make a judgment call based on where comparable bonds are trading. None of these options is satisfying, and the aggregate error across a large book can be material.
The desk-side experience of this problem is what Nikhil Krishnan brought when founding Deep MM in late 2023. The observation from that experience was that modern ML methods, applied to the full set of market signals that are actually available in the data, should be able to produce fair value estimates that are substantially more accurate than interpolation-based approaches. The hypothesis is testable, and the testing is what we are currently doing.
What the Signal Is
The Deep MM fair value signal is an OAS estimate for each bond in our coverage universe, updated on a five-minute schedule during market hours. OAS is the right output measure for corporate bonds because it strips out the duration component (the risk-free rate sensitivity) and isolates the spread that represents the bond's specific credit risk premium. Two bonds with the same OAS, the same duration, and the same rating should theoretically be worth the same price. When they diverge, one of them is mispriced relative to the other, and the signal tells you which direction the market is implying that divergence should close.
The signal is a point estimate accompanied by a confidence score and an uncertainty band. We want to be explicit that the confidence score is an internally calibrated measure of model certainty for that specific CUSIP at that time, not a guarantee of accuracy. For bonds with recent trade prints and deep peer groups, the confidence score is higher and the uncertainty band is narrower. For bonds that have been silent for weeks and have few comparable issues, the confidence score is lower and the band is wider.
The Data Inputs We Selected
The choice of input features was a significant portion of the early model development work. There are dozens of potential input signals relevant to corporate bond pricing, and not all of them carry reliable predictive content when subjected to proper out-of-sample testing. Here we describe the three main input families we have settled on.
Trade print data from TRACE: Every reported corporate bond trade through the FINRA TRACE system. This is the most direct evidence of where bonds are trading. For bonds with recent prints, this anchors the fair value estimate. For bonds without recent prints, the print history still provides calibration for normal spread levels and typical spread-to-peer relationships. We have data going back to 2019 for our initial training and validation work.
Dealer axis data: Dealer inventory axes indicate directional interest in specific credits. An axis showing a dealer is heavy in a bond (wants to sell) differs from an axis showing the dealer is covered (wants to buy). These axes are not widely aggregated or distributed, but they carry directional information about where institutional dealers think a bond should trade relative to current observable prices. Our system ingests and weights dealer axes as a supplementary signal, with explicit decay weighting based on the age of the axis update.
Cross-asset macro signals: The CDX investment grade and high yield indices, individual-name CDS spreads where active markets exist, Treasury rates at benchmark tenors, interest rate swap rates, and investment grade and high yield ETF flows. These signals carry information about the macro environment that drives corporate spreads. When credit spreads are widening broadly, that information should flow through to fair value estimates for individual bonds that may not have printed yet, rather than waiting for the next trade on each bond to reflect the market move.
Accuracy Targets We Are Building Toward
We want to be careful here. The numbers in this section are targets and early-stage development estimates, not validated production performance metrics. They represent where we believe the model should be able to perform if the development proceeds as we expect, based on our initial backtesting work and the signal quality we have observed in early testing.
For liquid IG corporate bonds (bonds with at least one TRACE print per week), our target is a median absolute OAS error below 10bp compared to subsequent observed prints. This is the regime where we expect the strongest performance because the data density supports the model directly.
For semi-liquid IG bonds (bonds printing roughly monthly), we are targeting median absolute error in the 15 to 25bp range. Performance here depends more on the cross-sectional and cross-asset features, as the direct trade print anchoring becomes weaker.
For illiquid IG bonds and liquid HY bonds, we are targeting median absolute error in the 25 to 50bp range. These are the segments where current alternatives (interpolation, gut checks, stale last prints) perform most poorly, and we believe this is where the model delivers the most practical value even at higher absolute error levels, because the baseline alternatives are substantially less accurate.
We intend to publish periodic validation notes as we gather more real-world performance data. The validation approach will use live observed trade prints as ground truth and compare them against the model's fair value estimate at the time of the trade. We will report the error distribution, not just the mean or median, so that a desk can assess what tails look like before relying on the signal for quoting decisions.
What We Are Not Claiming
We are not claiming that the Deep MM signal eliminates the need for desk judgment in corporate bond pricing. A model output is an input to human decision-making, not a replacement for it. The signal should help a trader anchor their price when the screen shows nothing, but the trader's knowledge of the specific credit situation, of current market tone, and of their own book's risk profile is essential context that the model does not have.
We are also not claiming that our approach is unique at the conceptual level. There are other firms working on model-derived fixed-income pricing. What we believe we are doing differently is focusing specifically on the illiquid tail of the corporate bond universe, where the data challenges are hardest and where the payoff to a working solution is greatest, rather than building primarily for the top few hundred benchmark names that anyone can price from market data.
The early access program we are running with a small number of institutional desks is designed to generate exactly the kind of real-world feedback that will tell us whether the signal performs as we expect and where it needs improvement. We treat those early partners as collaborators in the development process, not just as customers consuming a finished product.
What Comes Next
The next several research notes we publish will go deeper on specific aspects of the methodology: how we handle stale quotes, how we calibrate the HY model differently from IG, and how the cross-asset signal layer behaves during market stress events. The goal is to give technical counterparts at institutional desks enough visibility into the approach to evaluate it on its merits rather than on marketing descriptions.
We are building in public to the extent a small team can, because trust in a pricing signal has to be earned through transparency about the method, not through claims about the outcome.