Remove dealer fixed effects from the auto-hedging logit and R² drops from 0.29 to 0.01. Moving spread or depth by one-SD changes the probability of a hedge by 1.5 to 1.8 percentage points. The authors open their abstract with that contrast: AH variation is "primarily driven by time-invariant, entity-specific characteristics", while liquidity has "a more modest yet statistically significant influence". I accept the reported liquidity coefficients. I object to labels such as "primarily" or "almost all". Dealer effects account for most of the fit, yet the full model gets only to 0.29. Roughly 71% of the variation remains unexplained by anything in it.
Panzarino and Perrella assembled sixteen years of MTS trade prints into a census. It gives a trader almost no answer to whether auto-hedging pays.
What counts as a hedge?
Auto-hedging (AH, the authors' term) means a dealer fires an offsetting trade immediately after someone hits its quote. Suppose the dealer sells to an aggressor at 101. Within a second, it lifts another dealer's offer in the same bond to flatten its position. Stale quotes supply the rationale: when news moves fair value before a dealer updates, holding the short costs money. In the authors' example, news takes fair value to 103 before a buyer hits the stale 101 ask. Buying back at an offer still resting at 101.3 limits the loss to 0.30 rather than 2.
The confidential data cover MTS Italy trades from January 2010 to April 2026. Each print records a timestamp to the second, ISIN, size, price, side and counterparty codes. To earn the flag, the dealer whose quote was hit must become the aggressor within one second, trading in the offsetting direction. The baseline also requires the same ISIN.
Flagged trades were a residual share before 2016, then reached 5 to 10% in 2016 to 2019. The share peaked near 40% at end-2024; the conclusions describe it as "around a third" in recent years. By mid-2025, more than 22 market makers exceeded EUR 50 mn of AH a month. Before 2016, only one or two did.
The logit covers January 2023 to April 2026 and asks whether a trade gets auto-hedged. It uses market-wide bid-ask spread and top-of-book depth from the latest 5-minute snapshot, alongside trade size and direction. Dealer, bond and day fixed effects complete the model. Wider spreads and thinner depth each reduce the probability of a hedge. After the Treasury began including AH in its evaluations, a post-2025 interaction shows primary dealers reducing AH more than other dealers.
Most of the fit comes from dealer identity
The baseline R² is 0.29. Removing bond effects leaves 0.28; removing dealer effects leaves 0.01.
The liquidity coefficients are much smaller. Widening the spread by one-SD, about 4 bp, reduces AH probability by 1.76 pp. Increasing depth by one-SD, about EUR 5.7 bn, raises it by 1.47 pp. Sell-initiated trades leave the dealer long and are 1.69 pp more likely to be hedged. That is about the size of either liquidity shock.
The authors' conclusion here is right, and clearly stated. AH "cannot be fully understood through market conditions alone, pointing instead to the relevance of persistent differences across market makers". They acknowledge that they cannot observe what sits inside the dealer effects. Technology, risk tolerance and inventory management practices are among their proposed explanations. A desk instruction to flatten every position would qualify. A fixed effect remains a named residual; if it captures that instruction, AH is a policy choice with no edge attached.
Inference needs care, too. We did not find which pseudo-R² the 0.29 denotes, and the paper does not report the number of dealer clusters. Figure 3 shows more than 22 dealers exceeding EUR 50 mn of monthly AH, suggesting a few dozen at most. The authors say significance survives one-way and heteroskedasticity-robust errors. Those alternatives do not repair a small cluster count.
The clock limits the flag
The four rules make sense: timing, the same dealer becoming aggressor, the offsetting direction and the same bond. The timestamp is the weakness. Prints recorded to the second cannot fully establish the order of two trades within that second, even with a one-second window. An independent decision to lift an offer in the same bond can also trigger the flag. The authors cannot observe intent. They present the absence of a need for strategy information as a feature.
We did not find a size requirement either. The authors define a hedge as something that may "mitigate or fully neutralise" the shock. A small lift following a EUR 20 mn hit therefore counts just as a full offset does. Extending the window to five seconds leaves the results materially unchanged. Removing the same-ISIN restriction still puts 80 to 90% of flagged hedges in the same bond. These checks address window length and instrument choice. They leave ordering within a second unresolved.
Cheap, among the hedges that happened
The cost evidence deserves a slower reading. Over January 2023 to April 2026, 43% of flagged hedges traded at the same price as the fill they offset. Most of the remainder were within 2 cents. The authors interpret the declining frequency of higher-cost hedges as evidence that cost discourages AH.
Only completed hedges enter that histogram. A trade left unhedged because the offer disappeared, or became too expensive, contributes nothing to it. Selection explains part of the shape. The logit supplies the real test because it includes every trade: there, a one-SD increase in spread reduces hedging probability by 1.76 pp.
A zero price difference is a gross figure. Fees are missing, as is the more important counterfactual: what happens if the dealer holds the position? In the authors' example, the hedge earns its keep only because fair value has risen to 103. They acknowledge the reverse case. With fair value unchanged at 100, the 0.30 would be pure loss.
The counterparty matters as well. Another dealer owns the stale 101.3 offer in that example. Selling at 101.3 when fair value is 103 costs that dealer 1.70. More than 22 dealers auto-hedge on the same platform, so some of the adverse selection AH avoids falls on the slowest quoter. We did not find a PnL or markout comparison between hedged and unhedged inventory. None of the three future research avenues the authors propose prices the hedge. The nearest asks how AH affects market liquidity and market quality.
After the Treasury's 2025 criterion
Day effects begin drifting lower in early 2025, when the Treasury introduced the AH criterion. The paper also cites MEF (2026) for a decline in same-ISIN AH since 2025. Because the baseline flag requires the same ISIN, moving hedges into other bonds would appear as a reduction in AH in the interaction.
The text says the cross-instrument results in Table A4 "remain consistent with the baseline specification". It does not specify whether that statement includes the primary-dealer interaction, or report a magnitude for that interaction.
We could not test any of this ourselves.
Testing the mechanism requires trade prints with counterparty codes linking a dealer's passive fill to its next aggressive trade. It also requires MTS order-book depth at the time of the hedge. We hold minute and daily bars. Treasury futures cannot reproduce a cash interdealer book.
Until someone marks hedged fills against unhedged ones, dealer by dealer, the 40% peak measures a habit. The authors have tracked it through sixteen years of prints and have yet to price it.