The Dutch quarter-hourly book leads hourly order flow even when the model's near-critical cascade estimates are set aside. Those estimates are much more exposed: their scale falls exactly where the paper's 0.95 spectral-radius penalty begins.

Jhabli, AlSkaif, Bennin, Tekinerdogan and Pennings examine level 3 submissions in the Dutch EPEX SPOT intraday continuous market. A session pairs one Hourly XBID contract with its four Quarter-Hourly constituents. According to the paper, those product families accounted for 86% of Dutch traded volume in 2024. Its sample covers January, April, July and October 2024 and contains 2,915 delivery contracts. For each session, the authors use the final five hours, restricted to the window when all five contracts are open.

They fit a bivariate state-dependent Hawkes process to each contract. A submission increases the rate of later submissions in both its own book and the other book; that effect fades with time. Baselines and decay kernels vary between Normal and Stressed regimes. The Liquidity Stress Indicator averages slow z-scores of spread, 10-minute mid volatility and absolute queue imbalance with equal weights. The book enters the Stressed regime when that measure reaches or exceeds 0.264, the pooled sample's 75th percentile. A three-exponential kernel performs best in the comparison. Its pooled median Wasserstein-1 distance to Exp(1), the residual measure the paper treats as comparable across kernel families, is 0.0651 versus 0.0963 for the power law. Its likelihood exceeds the power law's in 99.5% of contracts and the two-exponential kernel's in 82.5%. The authors caution that the truncated power-law likelihood is not strictly comparable with the exponential likelihoods.

One finding concerns endogeneity. The median spectral radius is 0.9518 in Normal conditions and 0.9555 under stress, which the authors interpret as roughly 95% endogenous order flow. They make no claim that stress increases it: the within-contract difference is +0.0025, and only 53.5% of contracts show an increase.

The more interesting finding concerns direction. G measures direct branching, or the orders a submission triggers in the first round. Ψ = (I−G)⁻¹ adds up the whole cascade started by one exogenous order. In the Normal regime, total Quarter-Hourly-to-Hourly transmission has a Ψ of 6.675, against 2.390 in the reverse direction. Under stress, the figures are 5.238 and 3.393. The asymmetry index A = (Ψ_H→Q − Ψ_Q→H)/(Ψ_H→Q + Ψ_Q→H) shifts from −0.442 to −0.175. Its within-contract change is +0.148 (95% CI [+0.126, +0.201]); every month's interval excludes zero, though July's only just, at +0.000. A rises in 58.1% of contracts.

Where does the quarter-hourly lead come from?

The near-critical diagonal amplifies the cascades, as the paper argues. Median direct H→Q branching is 0.021, while median total H→Q transmission is 2.39. The per-contract ratio has a median of 113. Each book amplifies again by roughly 1/(1−0.9).

Yet the directional result survives without that amplification. In a 2×2 system, the off-diagonal entries of (I−G)⁻¹ have the same determinant. For each contract, Ψ_Q→H/Ψ_H→Q therefore equals G_Q→H/G_H→Q exactly. First-round coefficients settle which way the lead runs. Median G values imply an A of about −0.42 Normal and −0.17 Stressed, near the reported −0.442 and −0.175. The small gap follows from taking a median of ratios rather than a ratio of medians.

The diagonal deserves more caution. Its penalty starts at ρ* = 0.95 with weight 10³; fitted median radii of 0.952 (Normal) and 0.956 (Stressed) lie just above it. How much of that 0.95 is learned from the data, and how much comes from the penalty, remains open. The paper says the constraint binds in some fits and says the penalty magnitude is reported, though we did not find that magnitude in the results. Raise G_QQ from 0.90 to 0.95 and 1/(1−G_QQ) roughly doubles, from about 10 to 20. The claim that one quarter-hour order breeds "6.68 Hourly orders" is model-implied. The directional lead carries more weight.

Across months, the range of A contracts from 0.361 to 0.047. The paper treats the −0.99 correlation between calm-period asymmetry and the stress-induced change as another view of this convergence: stress moves each month toward a common configuration. That reading is fair, although the correlation supplies no further evidence. Stressed A lies between −0.16 and −0.21 every month, making the change approximately a constant minus the Normal level. Only approximately: the reported changes are within-contract medians (January +0.219), whereas subtracting monthly medians gives a different result (January −0.165 minus −0.541 is +0.376). The correlation also has just four months behind it.

Seconds matter here

Under stress, Q→H mean lag drops from 6.05 s to 0.93 s, and t90 from 17.8 s to 2.38 s. H→Q starts fast, with a t50 of 0.098 s. Its mean lag rises from 0.24 s to 0.50 s under stress.

I put less weight on the 6.5x timing figure than on the asymmetry. The paper says Tables XI to XV, including the timing table, depend only on the integrated kernel rather than on the allocation of mass across components. For G and Ψ, that holds. Mean lag uses the ratio of Σα/β² to Σα/β and changes when that allocation changes. The allocation is poorly pinned down: in the Normal regime, slow cross-product α reaches its lower bound in 48.3% of H→Q contracts and 49.4% of Q→H contracts. The Q→H slow half-life spans an interquartile range of 3.8 decades. The authors also link the fast component's position at the 1,000/s cap in 12 to 15% of contracts across the two cross-product channels to millisecond timestamps. A speed-up is plausible; its estimated size is less secure.

We could not test it. We have no Dutch intraday power data, and our finest bars are one minute. Under stress, a Q→H response completing 90% of its excitation in 2.38 s fits wholly inside one bar.

Hedging remains an untested implication

The paper warns that "a hedge calibrated on the leadership structure of calm periods becomes less reliable precisely when it is most needed." We did not find a hedging test, a trading test or transaction costs supporting that warning.

Serial dependence remains in most residual streams. The selected kernel passes Ljung-Box for 26.8% of contract streams, and the authors acknowledge that no kernel removes the dependence. Regimes are descriptive because the stress label comes from the same book the model seeks to explain. Still, the pooled threshold travels within the sample: the leave-one-session-out stressed fraction ranges from 0.181 to 0.316. The model covers submissions alone, since cancel-and-replace messages share millisecond timestamps.

The threshold's leave-one-session-out calibration is the paper's only out-of-sample check. A predictive test would change my view: after a live LSI crossing, do hourly submissions begin answering quarter-hour submissions within about a second, and does that result hold in a model with cancellations? Live use would also require a causal threshold. The current 0.264 is a percentile calculated from pooled in-sample data.