The academic literature on bond market microstructure is extensive. There is less applied work on which microstructure features actually generalize under ML methods in production settings, as opposed to in carefully constructed research backtests. This note surveys what we have found useful at Deep MM in terms of feature families, the conditioning variables that determine when those features carry signal, and the failure modes that emerged during development.
Why Microstructure Features for Bond Pricing
The standard approach to corporate bond fair value is fundamentally a cross-sectional problem: given a set of observable bond characteristics (duration, rating, sector, seniority) and current market conditions (Treasury rates, credit spreads, cross-asset signals), what is the appropriate OAS? This is a reasonable framework and it is the core of our pricing model.
Microstructure features add a time-series dimension to that cross-sectional picture. The question shifts from "what should this bond be worth given its characteristics" to "how is this bond's realized trading behavior informing us about current market conditions for this specific credit." These are different questions and they carry different but complementary information.
In practice, microstructure features matter most for two use cases: early detection of spread moves before they are fully reflected in sparse trade prints, and distinguishing between bonds in the same sector that are trading at different effective liquidity premiums despite similar fundamentals.
Trade Print Sequences: Volume Pattern and Direction
TRACE data provides the raw material here. A single trade print on a bond tells you price and size. A sequence of trade prints on the same bond tells you something about the direction of flow: is this credit attracting net buyer interest or net seller pressure?
The features we have found most useful from print sequences are not the prints themselves, but derived characteristics. The ratio of buy-initiated to sell-initiated trades over a rolling 20-day window (inferred from trade direction heuristics applied to the print sequence) carries directional signal for sub-30-day spread changes in our internal backtesting work (2023-2025 validation window). Consecutive same-direction prints, particularly when they cross typical daily volume thresholds, often precede further spread movement in the same direction.
The important caveat is that TRACE-based direction inference is noisy. We do not have the full order book. Customer-dealer direction assignments based on timing and price heuristics carry meaningful error rates, particularly for smaller transactions where the distinction between dealer-to-customer and customer-to-dealer flows is harder to determine. We treat these features as directional indicators with explicit uncertainty, not precise flow measurements.
Inter-Dealer Spread Dynamics
The spread between where an individual bond trades and where its closest peer bonds are trading at the same time carries information that static cross-sectional models miss. If Bond A and Bond B are two subordinated notes from the same financial issuer with similar duration and rating, they should trade at close to the same OAS. When they diverge by more than their historical distribution would predict, that divergence has a cause.
Building features around this spread-to-peer differential requires careful construction of peer groups. We define peer groups dynamically based on ratings bucket, sector, tenor bucket, seniority, and index membership. For each bond in the group, the deviation from the group-weighted median OAS is a feature. Persistent deviation in one direction (bond consistently trading wider than its peer median, for example) is a signal of either bond-specific credit concern or a liquidity premium that the cross-sectional model is not capturing.
In our experience, persistent positive deviation from peer OAS (trading wider) is more informative than negative deviation (trading tighter). Bonds that consistently trade tight to peers often do so because of index demand or specific investor demand, which is a market-structure phenomenon that may not persist. Bonds that consistently trade wide to peers in the same ratings bucket are usually doing so for a reason, and that reason tends to compound over shorter horizons.
Order Imbalance Proxies from TRACE
True order imbalance is not observable in TRACE data, but proxies can be constructed. The standard approach in the microstructure literature is to classify each TRACE print as customer-buy or customer-sell using a variant of the Lee-Ready algorithm adapted for OTC markets. Cumulative order flow imbalance over rolling windows then serves as a directional indicator.
We tested several imbalance proxy constructions during model development. The best-performing version in our validation work used a combination of: print-level price-to-dealer-quote deviation (where quote data was available through supplementary sources), transaction size relative to the bond's typical transaction size distribution, and time-of-day patterns within session. The resulting imbalance proxy carried predictive content for intra-day spread moves in liquid IG bonds, but performance degraded significantly for bonds printing fewer than three times per week.
This is a fundamental constraint on microstructure feature usefulness for the typical corporate bond. If a bond prints twice a week, there is not enough data to construct a reliable imbalance proxy at sub-day granularity. For this reason, we apply microstructure features selectively: they are active in the model for bonds above a print-frequency threshold and are replaced by cross-sectional features for infrequently trading bonds. The threshold currently sits at five prints per two-week rolling window for IG and three prints per two-week window for HY, based on stability analysis of feature importance across the frequency spectrum in internal backtesting.
Features That Did Not Generalize
It is worth describing the features that looked promising in initial backtesting but did not hold up under out-of-sample validation or in live model operation.
Intra-day volume weighting of spread levels showed strong in-sample predictive content but suffered significant look-ahead bias in our early implementations, where end-of-session TRACE data was being used to construct features for earlier-in-session observations. After correcting for this, the signal largely disappeared. Tick-by-tick price momentum features similarly looked good in research but produced excessive trading noise when applied to the sparse print environment of typical off-the-run corporates.
Cross-bond lead-lag relationships were another area we investigated. The hypothesis was that liquid bonds in a sector reprice first, and illiquid bonds in the same sector lag by some predictable interval. We found evidence of this pattern in sectors with clear on-the-run versus off-the-run structures, but the lag interval was unstable across market regimes and the signal was not strong enough to justify the complexity. We shelved it, though we revisit it periodically as our data history lengthens.
Generalizing Across Market Regimes
The most consistent finding from our microstructure feature development is that features built on absolute levels (absolute OAS, absolute transaction volumes) generalize poorly across market regimes. Features built on relative measures or deviations from recent history generalize much better.
A bond's OAS of 180bp means something different in a narrow-spread environment than in a wide-spread environment. A bond's print volume of 10 trades in a day means something different in high-activity periods than in low-activity periods. Features that capture deviation from the bond's own recent history, or deviation from its peer group's concurrent behavior, are more stable across the market regimes we have observed than those that embed level assumptions.
This is not a surprising conclusion. It is consistent with how experienced credit traders think about bond price behavior. What was somewhat unexpected was how sharply the generalization performance diverged between absolute and relative feature constructions in our formal out-of-sample analysis. The gap was larger than our prior would have suggested, which reinforced our commitment to relative feature construction as a design principle throughout the model.