Dealer axes are one of the more information-dense signals in the corporate bond market, and one of the least systematically used. Each axis represents a real-money statement from a dealer about where they are willing to transact on a specific bond. Collectively, the axes active on a given day across a dealer network create a picture of inventory positioning that has genuine forward-looking information content. We built the axis signal layer of Deep MM to extract and aggregate that content in a way that is useful for real-time fair value estimation.
What a Dealer Axis Actually Contains
A dealer axis is, at its core, an inventory management communication. When a dealer shows a bid axis on a specific CUSIP, they want to reduce a long position or establish short-cover inventory. When they show an offer axis, they are sourcing bonds they believe a buy-side client will want. The spread at which the axis is priced reflects their internal fair value assessment, adjusted for the size of the position and the urgency of their inventory need.
The axis carries three separable pieces of information that matter for fair value estimation: first, the spread level at which the dealer is willing to transact, which is an implicit statement about where they believe fair value lies; second, the direction (bid or offer), which signals whether dealer inventory is long or short in that name; third, the size, which reflects how urgently the dealer needs to move the position.
The limitation is that no single participant in the market sees axes from all dealers simultaneously. A buy-side desk typically receives axes from their panel of dealers, which may be 5 to 15 relationships. A dealer see their own inventory plus a limited view of competitor pricing through inter-dealer activity. Our model attempts to aggregate across what we receive to create a broader picture than any individual market participant has access to.
The Signal Extraction Problem
Raw axes are not directly useful as a pricing signal without processing. Several structural features of how axes are communicated and timed create noise that has to be removed before the signal content is accessible.
Stale Axis Decay
An axis published at 9:15am in a market that moves 10bp intraday is not the same signal at 2:30pm. We apply a time-decay function to each axis observation based on how much the broader credit market has moved since the axis was posted. When CDX IG has moved more than 2bp since the axis timestamp, we discount the axis spread accordingly, using the beta of that specific bond to the index over a trailing calibration window. The decay function is not linear with time; it is a function of both elapsed time and realized market movement during that window.
Inventory Direction Aggregation
When we observe bid axes from multiple dealers on the same CUSIP within a short window, the aggregate directional signal is stronger than any single axis. Two or three dealers simultaneously showing bids in the same name suggests a broader inventory positioning theme, not an idiosyncratic single-dealer position. We track directional concordance across axis observations and weight the aggregate signal accordingly. A single dealer showing a bid might reflect their own book. Three dealers simultaneously bidding the same bond typically reflects something broader: either a real-money seller is working the street, or multiple dealers formed a similar view of where the credit is going.
Size Normalization
A $1 million axis and a $25 million axis on the same CUSIP at the same spread do not carry the same informational weight. Large size axes from major dealers have historically been more reliable as fair value indicators than small soft axes, based on our internal analysis of axis-to-print outcomes. We normalize axis weights by size, dealer market-making activity in that sector, and historical accuracy of that dealer's axes in that rating cohort. The normalization is recalibrated periodically as new print data accumulates.
Incorporating Axes into the Fair Value Estimate
The axis signal enters the Deep MM ensemble as a constrained range rather than a point estimate. An active bid axis at OAS of 180bp on a bond our cross-sectional model estimates at 175bp tells us the dealer is willing to buy at 180bp, which means the true fair value is probably at or above 180bp from their perspective. This creates a directional constraint on our estimate rather than a competing point estimate.
When axis signals and cross-sectional signals agree within a reasonable tolerance (in our internal calibration work, roughly 5bp for IG bonds with moderate liquidity), the fair value estimate tightens significantly. Both signal sources are pointing to the same region of spread space, which increases confidence. When they diverge by more than that, we investigate whether the divergence reflects an information gap in the comparables (the cross-sectional model hasn't yet incorporated something the dealer knows) or an anomalous axis that should be down-weighted.
Axes as Leading Indicators
One of the more interesting patterns we identified in our internal validation is that dealer axes have forward predictive content relative to subsequent TRACE prints. When we see directional axis clustering in a specific bond (multiple dealers showing the same side), the probability that the next print will occur at a wider spread (in the case of offer-side clustering) or tighter spread (bid-side clustering) is noticeably higher than the base rate, based on our backtesting over the internal evaluation window.
This is consistent with market microstructure theory: dealer axes aggregate real-money order flow intention before it has crystallized into a transaction. An asset manager working a large program sell into the street will generate multiple concurrent bid axes from dealers competing for the flow. The bond will eventually print, probably at or near those bid levels. Our model captures the pre-print signal from the axis pattern rather than waiting for the print.
We are not claiming axes are uniformly reliable leading indicators. There are axes that reflect dealer marketing, not genuine inventory pressure, and those contain much less forward information. Distinguishing between real inventory axes and marketing axes is a classification problem we address through historical accuracy modeling at the dealer-sector level.
The Aggregate Network View
The individual axis signal is informative. The aggregate view across all axes we receive is more powerful still. On any given trading day, we observe axes across a meaningful portion of our covered CUSIP universe. Mapping the directional distribution of those axes reveals sector-level inventory positioning that is not visible from any individual market participant's vantage point.
A pattern where dealer bid axes are concentrated in one sector while offer axes cluster in another is a macro inventory signal. It tells us dealers are net long one sector and net short another, which has implications for fair value across both sectors: the net-long sector faces headwinds as dealers eventually work off their inventory, and the net-short sector may be better supported as dealers need to source bonds.
We incorporate this sector-level inventory signal as a tilt to the cross-sectional comparables baseline. A bond in a sector with clear dealer net-long positioning gets a modest spread widening adjustment relative to where pure comparables would place it. A bond in a dealer net-short sector gets the reverse. The tilt is small in magnitude but directionally consistent with what eventually happens when the inventory imbalance resolves.
Limitations and What We Do Not Claim
The axis signal has real limitations that are worth naming. Our model only sees axes from the sources available to us, which is not the full market. For any given CUSIP, we may be working with a partial picture of dealer positioning, and that partial picture could be directionally misleading if the dealers we observe are clustered on one side while others we do not see are on the other.
Axes in stressed market conditions behave differently than in normal conditions. When bid-ask spreads widen and dealers pull back from market making, the axes that remain visible are typically from dealers still willing to make markets, which is a selected subset. The axis signal degrades somewhat in stressed conditions, and our confidence intervals widen to reflect that. For the question of what axes contribute to fair value estimation, the answer is substantial value in normal conditions, meaningful but noisier value during moderate stress, and constrained value during acute dislocation when most dealers step back entirely.