How Deep MM prices a bond no one has traded in weeks
A technical overview of our data pipeline, model architecture, and calibration process.
Explore the methodologyThree signal families. Zero manual input.
Every Deep MM fair value is derived entirely from observable market data. No analyst overlays, no subjective adjustments.
TRACE post-trade data forms the primary signal. We ingest every reported print for covered CUSIPs in real time, weighting by recency, volume, and counterparty type. Even bonds with infrequent prints carry latent information in the direction and trend of their last several trades.
Real-time dealer inventory axes carry directional information that is invisible in trade data. A dealer who is long a stressed credit and axed to sell provides a different signal than one whose axis is neutral. We aggregate this across major dealers to extract net positioning signals.
CDS single-name spreads, swap rates, ETF flows, and index composition changes all carry information about where a bond should be priced relative to the broader market. These signals are most powerful for bonds with recent trade data gaps, providing the interpolating structure the model relies on.
Five stages from raw data to delivered signal
The Deep MM pipeline processes all three data families continuously, producing a refreshed fair value OAS for every covered CUSIP every five minutes.
Backtesting against subsequent prints
The primary validation methodology compares model-derived fair values against the next actual transaction print for each CUSIP. Because TRACE provides ground truth with a short delay, we can evaluate model accuracy on a rolling basis without survivorship bias.
We segment validation results by liquidity tier: bonds with weekly or better print frequency, bonds with monthly print frequency, and bonds that are nearly silent. Each tier has a distinct accuracy profile. We report separately so users understand what confidence to attach in each regime.
What 12bp mean error means in practice
The 12bp median fair-value error from our internal beta program is measured against subsequent transaction prints. For a bond trading at a 150bp OAS, this represents an 8% error in spread terms, which is within the bid-ask spread on all but the most liquid IG names.
Based on our internal backtesting across the 23-month validation window, stale last-trade pricing for a bond that has not printed in two weeks typically shows errors of 20-40bp in spread-widening environments and 15-25bp in rally scenarios. The model outperforms stale interpolation across all tested regimes.
Model accuracy across liquidity tiers
Read the full research behind the model
Our research papers cover signal construction, feature engineering decisions, and backtesting results in detail.