The canonical view of corporate bond pricing assumes there is a recent trade to anchor against. For the on-the-run issues in a major issuer's stack, that assumption holds reasonably well. For the remaining 70 to 75 percent of the investable IG universe by CUSIP count, it does not. Most investment grade bonds are genuinely thinly traded, and the pricing approaches commonly used on desks were built for a market with far more daily prints than the post-Volcker, post-MiFID reality produces.
The IG Illiquidity Profile
The common mental model of bond market liquidity conflates the most active names with the overall universe. A handful of benchmark issues from the largest IG issuers, particularly in financials and investment-grade utilities, do trade with high frequency. But TRACE data shows that a large fraction of CUSIPs in the USD IG universe trade fewer than five times a week, and many trade fewer than five times a month.
Several structural factors drive this. Bonds age off dealer flow quickly after issuance. The on-the-run new issue distributes through the primary market, and once dealers clear their allocation, secondary trading thins out. Bonds with smaller outstanding float attract less dealer attention for market making. And the Volcker Rule's constraints on proprietary inventory holding reduced dealer willingness to hold off-the-run positions, which means fewer dealers are actively making two-way markets in most names at any given time.
The result is a market where the last observable trade for a specific CUSIP may be two, three, or more weeks old. That last print carries a timestamp but no warranty: the spread environment may have moved since then, the issuer may have had a material announcement, and the dealers quoting that bond today may be at a very different level than where the last print cleared.
How Last-Print Interpolation Fails
Last-print interpolation is the most common fallback for pricing a bond that has not traded recently. The approach takes the last known spread, adjusts it proportionally based on sector or benchmark moves since then, and produces an estimated current level. It is intuitive, fast, and transparent. It is also systematically wrong in ways that compound over time.
Systematic Spread Drift
When the whole sector has moved since the last print, interpolating from the old spread produces a level that is too tight or too wide for the current environment. A proportional adjustment based on a sector index helps but does not fully capture the dynamics of individual bonds moving differently within the sector. Sector spread indices are dominated by the active, frequently traded names. A thinly traded bond with its own supply-demand dynamics may have diverged from its sector median without any print to record the divergence.
Issuer-Specific Events
Earnings misses, credit watch placements, leverage ratio changes, and M&A announcements all shift the credit curve for the affected issuer. If the last print predates the event, interpolation produces an estimate that has nothing to do with the current credit quality assessment. The error here is not a measurement error. It is a structural miss: the interpolation is estimating a spread for the pre-event entity, not the current one.
Bid-Offer Framing
When a trader quotes around a stale last print, the two-sided market they create is centered on a price that may be weeks old. The bid-offer may be appropriately sized for the perceived risk, but the midpoint anchor itself is wrong. Any counterparty with a better view of current fair value will consistently take the right side of that market. Over many transactions across a large book of thinly traded positions, this systematic mispricing is not neutral to PnL.
A Signal Architecture for Quiet Bonds
The Deep MM model treats fair value estimation for a thinly traded bond as a structured inference problem rather than a time-series extrapolation problem. Instead of anchoring on a single stale number, the model simultaneously draws from three signal families.
The first is cross-sectional comparables. Other bonds in the same issuer sector and rating cohort that have traded recently contribute to a peer-spread estimate. We weight peer contributions by rating proximity, modified duration match, and recent trade frequency. A BBB+ industrial with 8-year duration is more informative for a BBB+ industrial with 7-year duration than it is for an A-rated financial.
The second is dealer axes. When a dealer shows a bid or offer axis on a specific CUSIP, that represents a directional statement about where they are willing to transact. Even when the axis does not result in a print, its presence constrains the probable fair value range. An offer axis tells us the dealer's internal assessment is at or above that level. A bid axis tells us they see value there.
The third is cross-asset signals. The single-name CDS spread for the same obligor, CDX index levels for the relevant sub-index, and sector equity performance provide a continuous read on credit risk that does not require a bond trade to occur. When a bond has not traded in ten days, the CDS market has continued pricing the same underlying credit risk throughout.
Combining the Signal Families
Each signal family contributes a weighted prediction to the ensemble output. The weighting reflects estimated reliability for the specific bond on the specific observation date. A bond with no trade in 30 days but an active CDS market and several comparable prints in its sector will weight the cross-asset and comparables signals heavily. A bond with a recent dealer axis showing genuine size will weight the axis signal more.
When a bond has had no trade in the prior five trading days, the cross-sectional comparables and cross-asset signals carry the majority of the ensemble weight. As trade frequency drops below one print per month, we also widen the model's output confidence interval. This is not a failure of the approach. It is an honest representation of the estimation uncertainty: the less information available, the wider the range of plausible values, and we convey that to the desk rather than suppressing it.
A concrete example from our internal validation work: a BBB+ rated utility issuer with a 10-year note outstanding from a 2022 issuance had been trading fewer than two times per week by late 2024, with some weeks showing no print at all. In our backtesting analysis using internal data, the last-print interpolation for that bond was anchoring approximately 3 to 4bp too tight relative to subsequent prints, based on comparison with trades that eventually occurred. Our model's estimates, incorporating the issuer's CDS spread movement and comparable-maturity utility peers, were within a tighter error band on average over that validation window. These numbers come from internal backtesting on a specific sub-sector and should be read as directional rather than as general accuracy benchmarks.
Where the Model Reaches Its Limits
Not every thinly traded bond is equally tractable. For bonds with unusual structural features, whether callable, putable, or convertible, the comparable selection problem is more difficult. The model's confidence intervals widen accordingly, and we are explicit about that in the output.
For bonds from issuers with no active single-name CDS market, the cross-asset signal leg is absent. The inference leans more heavily on peer-group comparables, which is workable but less precise. Small-cap industrials and some niche financial issuers with no CDS reference fall into this category.
We are not claiming this approach solves the fundamental illiquidity problem. A bond with no reference CDS, a thin peer group, and an idiosyncratic credit story is genuinely hard to price. What the model provides is a structured, auditable estimate with an explicit confidence score attached. That is more actionable for a risk function than an unstated assumption that the last print, wherever and whenever it occurred, is still a valid anchor. The assumption is almost always wrong by the time three weeks have passed. The model is wrong too, but in a bounded, quantifiable way.