Woo attributes 41% of Samsung Electronics' closing displacement variance to the SK Hynix leveraged complex trading next to it in the same Seoul closing auction, on his conservative calibration. Samsung's own loop gain is 0.24, a multiplier of 1.32, which any product-level dashboard would wave through. That diagonal is imported from Zhao's calibration. Whatever you think of the identification, the gap between what a receiver's own gain says and what its close is doing is the finding.
What is actually being estimated?
A fund promising L times the daily return has to trade A(L^2 - L)r into the close, where A is fund assets and r the day's move. The coefficient is 2 for a +2x fund and 6 for a -2x fund, so long and inverse products push the same way rather than offsetting. Zhao, whose calibration Woo borrows, prices this self-reference with a scalar loop gain: closing-venue impact times the complex's rebalancing capital. Spillovers to correlated neighbours enter that setup as bias to be signed.
Woo promotes the spillover to the object of study. He writes n coupled complexes as a within-close fixed point, r2 = (I - L)^{-1}(L r1 + v), with L = Phi diag(gamma). Phi holds own- and cross-impact coefficients; gamma_i is complex i's rebalancing capital divided by average traded value. Two consequences follow. Cycle amplification: for nonnegative L the spectral radius is at least the largest own gain, strictly larger only when coupling runs both ways. Transmitted displacement: the off-diagonal of (I - L)^{-1} - I is positive under one-way coupling alone, sized by the sender's capital, and leaves the receiver's spectral radius untouched.
The part a practitioner can run is the estimator. Displacement reverts overnight, so regressing close-to-next-open returns on the vector of same-day returns recovers the displacement-response matrix up to an overnight correction share theta. Inverting that gives L. Prices and public fund filings only; no signed order flow.
Simulation next. In the two-asset design (T = 250, 500 paths) the spectral radius comes back with RMSE 0.005. One configuration is calibrated to the blind spot: own gains of 0.35 each, coupling of 0.30, so rho(L) = 0.65 against an "unsafe" threshold of 0.5. The scalar monitor clears every path as safe. The matrix monitor flags every path as unsafe.
Empirically he runs a difference-in-differences of the cross overnight reversal coefficient around the 27 May 2026 Korean single-stock LETF launch: 16 stocks, roughly 95 pre-launch and 56 to 60 post-launch trading days, 20 funds from KRX daily statistics. He then repeats the design on the U.S. MSTR/BITO/COIN complexes with CRSP daily data, January 2023 to December 2024, split at the 15 August 2024 MSTX launch. There is no strategy here and no performance accounting of any kind; it is a monitoring paper.
We could not test any of this. The Korean episode is only reachable for us through U.S.-listed leveraged ETFs and their U.S. equity or crypto underlyings, and the Korean point estimates and calibrations do not transfer to that market. The mechanism could in principle survive that swap, since the mandated close trade is the same arithmetic on any venue, and Woo's own U.S. panel is built exactly that way. The blocking piece is more basic than geography. Building K = sum_f A_f(L_f^2 - L_f) needs point-in-time daily assets and leverage multiples for every product in a complex, which we do not hold.
The Korean evidence runs one direction
The Hynix to Samsung overnight cross-coefficient goes from continuation before launch (t = +3.2) to reversal after (beta = -0.195, t = -1.1), DiD z = -2.82. It survives dropping the 19 August U.S.-shock day (z = -2.73) and Newey-West with five lags (z = -2.72). Because only two stocks are treated, inference rests on exact randomization over all 182 ordered pairs of fourteen non-treated large caps: the treated statistic ranks 0/182, one-sided p = 0.0055, two-sided 0.011. A ten-day block bootstrap puts the DiD mean at -0.68 with a 95% interval of [-1.25, -0.15].
Two details make this more than a cross-correlation. Interacting the Hynix return with standardized gamma gives beta = -0.093, t = -2.21, so the loading tracks the sender's disclosed rebalancing capital. And the effect only appears in the direction relative size predicts: the Hynix complex peaked near KRW 37.5tn against Samsung's 6.6tn, about 5.7x, and the reverse direction is undetectable. Woo also flags that without the same-calendar-day SMH control the treated coefficients come out positive and the design would reject coupling outright, with U.S. loadings entering at t around 5. His instruction to anyone running overnight-reversal designs on Asian markets is that this control is mandatory.
He records the objections himself. Pre-launch loading is strongly positive, so a negative DiD could be a decaying lead-lag rather than arriving coupling. He says plainly that 56 days cannot adjudicate it, and reports both a conservative and an upper-bound basis instead. The treated pair's pre-period correlation is 0.86 against a control median of 0.55, which he calls an imbalance. His high-correlation subset keeps pairs with pre-period correlation of at least 0.3, which is 178 of the 182 pairs, and returns p = 0.0056. My reading is that this leaves the randomization benchmark imbalanced on the covariate that matters most. One cell runs the wrong way: Samsung to Hynix drifts positive in the next-day window, z up to +2.16, unexplained.
No two-way coupling in either panel
Under the M-map, the assembled Korean matrix gives rho(L-hat) between 0.612 and 0.644 against a scalar maximum of 0.61. An excess of +0.002 to +0.034, on that same conservative mapping. Woo's own line is the right one: "the honest summary is that Korean cycle amplification is statistically indistinguishable from zero, by the structure of the episode rather than by noise," because the detected coupling is one-directional and the matrix is quasi-triangular. Note where the binding stability statement in the paper actually comes from. The dynamic bound is rho(L) < 1 - theta/2 = 0.56 at theta = 0.88, and SK Hynix's scalar gain of 0.61 already violates it. A scalar did that.
So the flagship case funds Proposition 2 and leaves Proposition 1 as a theorem with a simulation attached. The U.S. panel was supposed to be the symmetric configuration where cycle amplification is largest. All six treated cross-reversal directions are null, the most negative at z = -1.45 for COIN to MSTR, with placebos inside 1.22. MSTR scaled capital reached 1.72 and averaged 0.77 post-launch. Woo reads that as coupling being graded by venue depth, consistent with U.S. scalar gains of 0.03 to 0.22 versus Korea's 0.61, and it does work as a scale placebo for the estimator. Neither of his two panels contains a measured two-way coupling product. He has an answer ready: the symmetric common-factor configuration is the one now growing in U.S. crypto-linked products, and his CRSP coverage stops in December 2024, before that growth. Any review that wants to press the point has to beat that reply.
The estimator's dependencies compound this. Theta = 0.88 and the diagonal own gains are imported from Zhao rather than jointly estimated. Woo tries to recover the diagonal himself and gets 0.42 for Hynix (z = -2.0) against the calibrated 0.61. Close enough to be reassuring. But the Samsung own-effect comes back wrong-signed with |z| < 0.5, so the calibrated diagonal is retained. The measurement convention does more damage. His regressor is the close-to-close day return, while the model calls for the pre-close return. Switching from the M-map to the L-map moves the conservative-basis imported share from about 0.41 to a range of 0.84 to 0.91. It also moves rho up to 0.644, and to 0.86 on the upper-bound basis. He declines to claim the two conventions bound the truth, and he is right to. The 41% carries a delta-method range of 30 to 84 percent on the post coefficient, and the share runs 0.35 to 0.91 across the theta and level/DiD variants. Read the sign and the order of magnitude.
One more fact deserves repeating to anyone who thinks about intervention design: the 31 July 2026 deposit rule cut secondary-market turnover by roughly 90% and left gamma of order one throughout the post-launch window.
One thing would change my mind on the cycle half. A panel deep enough to show a two-way coupling product, in symmetric complexes written on one factor. Then rho(L-hat) clearing the 0.56 dynamic bound while every own gain sits under it.
Until then, the part of this framework I would put on a screen prices the displacement one complex imports from its neighbour: the off-diagonal transmission statistic. The sign restriction is what keeps it from firing on ordinary cross-autocorrelation. Woo's naive unsigned comparison false-alarms on 100% of no-coupling paths, against 0.04 to 0.05 for the controlled version at identical power.