Technical papers from the Deep MM team covering OAS fair value modeling, liquidity regimes, dealer axis signals, and machine learning applied to fixed-income markets.
Direct comparison of deep learning inference against interpolation methodologies for investment grade corporate bonds with infrequent trade prints. Cross-validated on 18 months of TRACE data across 1,200 CUSIPs with monthly or lower print frequency.
Read Paper
Empirical study of the economic cost of marking illiquid bond positions at last trade versus a continuously updated model fair value, using realized PnL outcomes from subsequent transactions.
Technical description of how Deep MM aggregates and conditions dealer inventory axis data to produce a directional pricing signal for off-the-run corporate bonds. Discusses signal decay and stale-axis handling.
Analysis of model behavior during three stress periods observed in the 2023-2025 dataset. Discusses how increased cross-asset weighting during spread regime shifts maintains signal accuracy when TRACE print frequency drops.
Feature importance analysis of CDS spreads, ETF flow, and swap rate signals in the Deep MM model ensemble. Demonstrates which cross-asset signals contribute the most predictive power and under what conditions.
Architecture decisions for building ML pricing models on corporate bond datasets where most CUSIPs have fewer than 10 observed trades in a rolling quarter. Discusses data augmentation, transfer learning from similar issuers, and confidence scoring under sparsity.
Empirical evidence that high yield and investment grade bonds operate in structurally different liquidity regimes, requiring separate model calibrations. Documents the feature distribution and accuracy differences when a single model is applied to both universes versus regime-separated models.
Case study analysis of how stale quotes form, propagate, and create systematic errors in corporate bond portfolios. Uses TRACE post-trade data to measure the economic cost of stale mark exposure across a sample IG portfolio during a spread-widening episode.
Survey of machine learning approaches applied to the specific characteristics of fixed-income market microstructure: trade-level data sparsity, dealer-intermediated price discovery, and the absence of centralized limit order books. Covers gradient boosting, deep learning, and hybrid approaches.
How the corporate bond pricing problem changes when applied to hard-currency EM corporates in frontier market jurisdictions. Discusses sovereign risk integration, thinner cross-asset signal availability, and model confidence degradation in data-sparse EM names.
Quantitative analysis of interpolation errors introduced when extracting benchmark spread from interpolated treasury or swap curves. Shows how small yield curve interpolation biases compound into meaningful OAS miscalculations for bonds far from on-the-run benchmarks.
The original research preview introducing the Deep MM methodology to the market. Covers the core hypothesis, data inputs, model structure, and initial validation results from the early beta program. Starting point for understanding the Deep MM approach.
The papers describe the methodology. The signal delivers it, updated every five minutes for 4,200+ CUSIPs.