The yen carry unwind barely registered in Diebold-Yilmaz total connectedness: 55.40 in July, 55.20 in August. Those readings cover the four dollar crosses. JPY rose 7.46% against the dollar in August 2024, its eighth largest monthly gain since 1975, while the alternative measure introduced by Li and Gong climbed from 60.13 to 63.06. These are the authors' own figures. Their abstract says "Connectedness rises sharply around funding stress", yet the standard DY index scarcely changed in the funding-stress month. Whether the network "saw" the unwind depends on the decomposition. That dependence is the paper's most useful result.
A shared covariance, different allocations
Li and Gong use monthly dollar returns on EUR, GBP, JPY and CHF, calculated from Bank of England monthly-average spot rates. The sample contains 612 observations from February 1975 to January 2026. Before 1999, the euro is synthetic and IMF-weighted. The authors filter a first-order time-varying VAR with forgetting factors of 0.99 for coefficients and 0.96 for covariance, then extract a 12-month forecast-error covariance at each date.
Both measures come from that covariance. Diebold-Yilmaz (DY) allocates each currency's forecast variance to individual, correlated innovations and rescales each row to sum to one. The paper's TVP-VAR-R2 measures how much of a currency's forecast error the other three explain jointly. Genizi's decomposition divides that R2 among them, giving nonnegative shares that add up exactly. Each method produces a total connectedness index (TCI), directional TO and FROM totals, and NET, defined as TO minus FROM, with a positive value indicating a net contributor. Both also produce net pairwise NPDC, the balance of transmission within each pair.
The aggregate agreement is close.
The TCIs have a correlation of 0.918. R2 is higher in 611 of 612 months, with an average gap of 9.92 points. About three quarters of the difference comes from EUR-CHF: that pair accounts for 29.49 of the 39.66-point increase in off-diagonal entries (74.34%). Average NET rankings agree exactly: EUR, CHF, GBP, JPY. Bilateral rankings diverge. DY puts EUR-GBP first, with mean absolute NPDC of 5.44; R2 puts JPY-CHF first at 2.67, compared with 2.38 for EUR-GBP.
What must an implementer choose?
The inputs are unusually well specified. The paper supplies the four Bank of England series codes, forgetting factors, forecast horizon and alternative priors. Its code includes the Monte Carlo simulations comparing the recursive estimator with 36-, 60- and 120-month rolling windows. Three implementation choices still need attention.
Start with the horizon. Every figure reported in Table 6 for H=6 and H=24 matches the 12-month baseline within 0.001 percentage points before rounding. According to the paper, VAR propagation enters at horizons of two months and longer. With identical figures at 6, 12 and 24 months, propagation contributes almost nothing. The "multi-period" R2 behaves like a decomposition of the innovation correlation matrix. Implementers can choose H freely. That freedom also tells us how little the VAR contributes.
The initialization uses the full sample. The authors say their baseline OLS initialization does so, "so the resulting historical paths are filtered rather than real-time estimates." They also centre returns before estimation, which we interpret as using a full-sample mean. A live signal would require a training-window prior and trailing centring. Those changes produce a different estimator.
Monthly averages are an intentional choice. They avoid non-synchronous closes across trading centres, a fair rationale. But there is only one observation per month. An unwind lasting days becomes a single data point, and the episode table averages across 25-month windows.
The least connected currency can still be a net receiver
NET records a balance and conceals the gross amounts. The abstract describes net receivers as "the currencies most affected by the rest of the portfolio." The reported receipts do not support that reading. JPY receives 35.39 under R2, the lowest of the four; its DY receipt of 37.58 is also the lowest. The other currencies therefore leave 64.61 of its forecast variance unexplained.
EUR receives 79.09 under R2 and 58.06 under DY. Its own share is just 20.91, even though its NET is +3.75. JPY earns the "receiver" label because it sends even less than it receives. For diversification, EUR and CHF share the most variance, with a pairwise Pearson correlation of 0.86. The negative-NET pair shares the least.
The recommendation that "GBP and JPY should be considered together when the funding currency moves sharply" thus depends on a label for transmission imbalance. The abstract describes GBP and JPY as "exposed to the unwind". The paper never tests that description against returns, hedge payoffs or losses. The authors acknowledge that exact currency and pairwise rankings "do not capture funding costs, hedge payoffs, or exposures elsewhere in the investor's portfolio."
Little changes in the August 2024 window
The authors defend the result by saying the large JPY return "does not show that connectedness was concentrated in JPY." Correct. Yet every period they examine would have delivered the same network answer. Across all seven 25-month windows, beginning with the U.S. Dollar Support Program and ending with the 2024 unwind, GBP and JPY occupy the lowest-NET positions under both measures. JPY is a net receiver under both in 97.39% of months. The abstract itself says August 2024 "shows the same pattern."
The episode adds nothing to the full-sample average on the question of which currencies receive. Its distinctive feature is narrower: the measures disagree over whether GBP or JPY ranks lowest.
Considerable uncertainty sits beneath that stable label. The correlation between JPY's monthly NET series is only 0.213. In the post-1999 sample, GBP has the lowest DY NET rather than JPY. Smoother dynamics (0.99, 0.99) put EUR-GBP first under both measures. We found formal inference only in the supplement: a HAC standard error and two block-bootstrap intervals for the average TCI gap. The main text supplies no intervals for the NET or NPDC rankings that shift across these rows.
Why we did not backtest the connectedness signal
We lack FX spot series for these crosses, and their futures roots are outside our tradable coverage. Our data also begin long after 1975. Even if we had the data, a causal implementation would need to replace the full-sample prior. We would then be testing our modification rather than the published baseline.
Watch the DY-R2 gap
The difference between the TCIs deserves attention. It averages 9.92, has a standard deviation of 4.47, and reached 31.08 in October 2008. During August 2024, the gap widened from 4.73 to 7.86 in a month while DY remained flat. A widening spread indicates tighter comovement in a form DY does not assign to individual innovations. It is a cheap monitoring flag, and its October 2008 peak shows that it spikes under stress.
The paper tests no P&L. Its own Monte Carlo gives the recursive estimator an average delay of 15.59 months in detecting a directional reversal. The authors note that this is close to the 36-month window and faster than longer windows. It remains far too slow for an unwind that unfolded within a single month.