A pricing model's performance in normal conditions is table stakes. What matters to a credit desk under pressure is how the model behaves when conditions turn abnormal: when spreads move 20bp in a session, when dealer liquidity vanishes, when the cross-asset correlations that held for two years break down in a single week. Every fair value model has latency built in. The question is how long it takes for that latency to become operationally dangerous, and what you do to shorten it.
The Regime Problem in Credit Pricing
Most statistical models of credit spreads are trained predominantly on normal-regime data because normal-regime data is most of what exists. Dislocation events are, by definition, infrequent. A model trained on 18 months of data from a benign credit environment will have learned relationships between variables that are specific to that environment. During a stress event, those relationships change. Spreads that were highly correlated with a specific risk factor may decorrelate; the OAS beta of single-name credits to the CDX index may compress or spike; bid-ask spreads widen, making even observed prints less reliable as fair value anchors.
A model that does not explicitly account for regime shifts will produce estimates in stress conditions that are systematically too tight, too slow to move, or anchored on a correlation structure that no longer applies. That is the problem the regime-detection layer in Deep MM is designed to address.
Defining a Widening Regime
We treat regime classification as a continuous, probabilistically expressed state rather than a binary normal-or-stress switch. At each model update cycle, we compute a regime score that reflects the current stress level across several observable market dimensions.
The primary inputs are: CDX IG 5-year spread level relative to its 60-day rolling mean; realized intraday spread volatility in the investment-grade index over the prior 5 trading days; the ratio of ask-side axes to bid-side axes across our observed dealer flow; and the VIX level as a cross-asset volatility proxy. These inputs are combined into a composite score that runs from 0 (clearly normal) to 1 (full dislocation).
The thresholds for regime transitions were calibrated using historical stress periods. We are explicit that the calibration window is finite and that the model has not been tested through every conceivable type of credit dislocation. The regime layer is designed to respond well to the stress patterns most similar to those in our training data, and more cautiously (wider confidence intervals, higher model uncertainty flags) when the stress pattern is novel.
How the Regime Score Changes the Signal
Recalibrated Cross-Asset Betas
During normal conditions, we estimate each bond's OAS beta to CDX IG using a trailing 60-day regression window. During stress, this beta tends to compress for high-quality credits (they widen less than the index) and spike for lower-quality IG credits (they widen more). When our regime score crosses a moderate stress threshold, we shift the beta estimation window to a shorter trailing period and apply a stress-period beta adjustment derived from historical dislocation patterns.
The practical effect is that when CDX IG widens 15bp over three days, our model re-estimates each bond's fair value using a spread beta that reflects how that type of bond actually behaves in spread-widening regimes, not the smooth-conditions beta it carried the prior week. For a BBB-rated industrial issuer with meaningful financial leverage, the stress beta is substantially higher than the calm-period beta. For an AA-rated utility with defensive cash flows, it is lower.
Reduced Weight on Stale Prints
In a fast-moving market, a print from two days ago is dramatically less informative than it would be in a stable market. During elevated regime scores, we accelerate the time-decay function applied to all trade print observations. A print that would carry substantial weight after a 3-day gap in normal conditions may carry very little weight after the same gap during an active widening event.
This prevents a common failure mode in print-anchored models: stubbornly holding onto a last print from before the stress onset as if it still represents fair value. In practice, when a credit event hits and the market reprices, the most recently traded bonds will provide new prints quickly. The thinly traded names will not. Without decay acceleration, the model treats the two categories symmetrically, which means the thinly traded bonds stay anchored on a now-stale pre-stress print far too long.
Cross-Asset Signal Elevation
Single-name CDS spreads trade continuously and reprice in real time during credit events, even when the bond market is thin or disorderly. When our regime score is elevated, we increase the weighting on the cross-asset signal leg, particularly the CDS spread for the same reference entity. This shifts the model from being primarily informed by the bond market (through prints and axes) toward being more informed by the derivatives market, which is carrying more current information during dislocation.
The crossover is not instantaneous and not complete. A CDS spread and a bond OAS are related but not identical: the difference between them is influenced by the CTD option value, funding, and basis dynamics. We maintain a calibrated CDS-to-OAS basis estimate that is updated using any bond prints that do occur, and we apply that basis when translating CDS-implied fair value into OAS terms.
Validation Against Historical Widening Episodes
We evaluated the regime-detection layer's performance against several historical spread-widening periods captured in our internal data history. For each episode, we compared the absolute error of our model's fair value estimates (measured against subsequent prints within a 5-day window) against a simple last-print interpolation baseline.
The regime layer showed the largest error-reduction advantage in the first 3 to 5 days of a spread-widening event, which is the period during which last-print models are most severely anchored on pre-stress data. As the stress period continued and new prints accumulated, both approaches began to converge because the print data itself updated. The value of the regime layer is therefore most concentrated in the early and fast-moving phase of a dislocation, which is also when pricing accuracy matters most for execution decisions.
These findings come from internal backtesting and should be read as directional, not as generalized benchmarks. The degree of outperformance depends heavily on how quickly the specific stress event moved and how much print data was available during it.
What the Regime Layer Does Not Solve
Regime adaptation is not a solution to the fundamental problem of extreme dislocation. When bid-ask spreads widen to 50bp or 100bp, when dealers step back from market making entirely, when the bond market is effectively shut for a category of credits, the model's output should be interpreted with extreme caution. We convey this through a maximum uncertainty flag that fires when the regime score is above a high threshold, at which point we recommend treating the fair value output as a directional reference rather than a tradable anchor.
There is also a feedback problem inherent in any live model during stress: when the market itself is thin and disorderly, the inputs to the regime score (axes, prints, CDS) become noisier. The regime detection that is most needed during severe stress is operating with the lowest-quality inputs. We acknowledge this openly. The design goal is not to produce accurate fair values in all conditions but to produce well-calibrated uncertainty estimates in all conditions, so the desk always knows how much weight to place on the signal they are receiving.