When we started building the Deep MM fair value signal, one of the first architectural temptations was to train a single liquidity-adjusted pricing model across the full corporate bond universe. There is a superficial logic to it: both IG and HY are corporate bonds with OAS as the natural spread measure, both trade in TRACE-reported OTC markets, and pooling data generally helps sample size. After working through the structural characteristics of each segment, we concluded that a shared model was the wrong approach. Here is the reasoning behind that decision.
The Trade Print Landscape Looks Similar But Is Not
At first glance, IG and HY bonds share the same surface-level data problem: most bonds trade infrequently. A large IG issuer might have 40 to 50 CUSIPs outstanding, and in any given week, half of those will show no TRACE prints at all. A mid-size HY issuer typically has fewer outstanding CUSIPs, but those bonds also print rarely. The scarcity of trade data is common to both segments.
The similarity ends there. In IG markets, liquidity is primarily a function of rates exposure, not credit uncertainty. When an IG bond goes quiet, it usually means rates are stable and investors are content to hold. The OAS on a 10-year industrial bond that has not printed in two weeks has probably not moved materially, and proxying its fair value from similarly positioned cousins with comparable duration and sector usually produces a reasonable estimate.
In HY markets, quiet periods carry different information. When a bond stops printing, it might mean rates are stable. It might also mean dealers are reluctant to commit capital to a credit they are uncertain about. In HY, the reason for the silence matters as much as the silence itself. A model that treats two-week staleness identically across IG and HY will systematically missprice HY bonds during elevated-risk periods.
Feature Engineering Diverges by Segment
The most consequential architectural difference is in what explanatory variables carry predictive weight.
For IG pricing, the dominant features in our model are maturity and duration profile, sector and ratings bucket, the corporate-Treasury spread basis at each tenor, CDS spreads for names with active single-name markets, and issuer-level recent print history normalized by issue count. Spreads for IG bonds are explained largely by macro rate sensitivity and broad sector-level credit risk. The idiosyncratic component is real, but it is smaller and slower-moving. An IG bond that has not printed in two weeks is typically well-predicted by its peer group.
HY pricing requires a different feature set. The issuer-specific credit story is the primary driver, and it cannot be proxied reliably by sector averages. The features that carry weight in our HY model include gross leverage and interest coverage from the most recent reporting period, capital structure position (senior secured versus senior unsecured versus subordinated), the nearest-term maturity cliff in the issuer's overall debt schedule, and sector-specific stress indicators. For energy HY, that means commodity price inputs. For retail HY, it means same-store sales proxies. For consumer cyclicals, consumer confidence and credit card delinquency data carry signal that simply does not matter for IG utilities.
None of these features are irrelevant for IG, but their signal-to-noise ratio is much lower in that segment. Pooling IG and HY in a single model forces a choice: either dilute the HY-specific features or overweight them relative to what IG bonds require, with the result being a model that produces mediocre estimates in both segments.
The Calibration Cadence Problem
IG fair value signals can be recalibrated on a weekly schedule without significant accuracy loss. Credit conditions in the IG market change gradually, and a model trained on the prior week's data is generally applicable to the current week unless there is a macro shock.
HY requires daily recalibration. Credit spreads in HY are more volatile, issuer news impacts prices with more force, and the relationship between features and fair value can shift materially within a week. In our internal backtesting work (historical data, 2023-2024 validation window), HY model error grows noticeably from day 2 onward when recalibration is skipped. IG model error under the same conditions remains relatively flat for five to seven days. Running the same calibration schedule for both segments imposes a false uniformity that costs accuracy in HY without meaningfully improving IG.
How the Failure Mode Looks in Practice
The failure mode of applying an IG-calibrated model to HY bonds is systematic underestimation of spread dispersion. IG models learn to cluster bonds within tight sector bands and rely heavily on duration as an explanatory variable. Applied to HY, this produces two predictable errors.
First, the model underestimates the range of fair values for HY bonds with the same duration. A BB-rated energy bond and a BB-rated consumer staples bond with identical modified duration can be 80 to 120bp apart in fair value. An IG model, trained to expect 20 to 30bp of intra-sector spread variation for equivalent duration, misses this dispersion entirely.
Second, IG models are slow to respond to issuer-specific news. The IG model relies on peer-group interpolation, which is a strength in low-volatility environments and a weakness when a single HY name needs rapid repricing after a covenant breach announcement or leverage guidance revision.
We ran this experiment during model development, applying our IG ensemble to a HY validation set and comparing the error distribution to our HY-specific model. The median absolute error was roughly 40% higher for the IG model on HY bonds (internal validation, 2024 holdout set). The gap was largest in the BB/B boundary segment and during months with elevated single-name credit events.
Our Architecture: Two Models, One Cross-Asset Layer
The current Deep MM production approach maintains separate IG and HY model instances with distinct feature sets, calibration schedules, and staleness threshold parameters. What we share across both is the cross-asset signal layer: CDX IG and CDX HY index levels, investment grade and high yield ETF flows, and Treasury and swap rate inputs. These macro signals carry information relevant to both segments, and computing them twice independently would produce redundancy without benefit.
The architecture is: cross-asset features computed once and fed into both models as common inputs, with segment-specific credit features processed separately and combined at the ensemble aggregation step. Model outputs for the same CUSIP are never averaged across the IG and HY instances. A bond is in one regime at a time, and the appropriate model instance handles it.
Crossover Credits and the Regime Boundary
We want to be precise about one nuance. We are not arguing that IG and HY are informationally isolated segments with no signal crossing the boundary. Rising-star bonds moving from HY to IG carry forward information from their HY pricing history during the transition period. Fallen angels moving the other direction need their pricing context reset, and the IG model's peer-group interpolation will actively mislead during the first weeks after a downgrade.
Crossover credits get specific handling in our system: for bonds within one notch of the IG/HY boundary, we run both model instances and flag the result with a higher uncertainty score. We do not average the two outputs. Instead, we treat the wider of the two spread estimates as the appropriate conservative bound and communicate that uncertainty explicitly in the signal output. This is a deliberate choice: during periods of ratings volatility, false precision is more dangerous than acknowledged uncertainty.