A desk should not trade the paper's 16 to 137 seconds as a price-drift horizon. That interval marks when half the extra price events have arrived after a macro release. Prices can absorb news on another clock. Dutta provides separate expected-drift curves, so the risk lies in reading the abstract's "how fast prices absorb news" as a directional trading signal. The 16 to 137 s figure counts threshold crossings.

Dutta's multivariate Hawkes model, MSX, represents each kernel as a nonnegative sum of Erlang densities on a logarithmic grid of decay rates. Its parameters enter intensity linearly, making the log-likelihood concave. Each fit carries a Fenchel duality gap, a computable bound on distance from the global optimum. For six assets, an up or down price event occurs when log price has moved by delta of 0.25 to 3bp since the previous event. The assets are EUR/USD, USD/JPY, gold, an S&P 500 CFD, and Binance BTC and ETH futures. The study draws on 670 releases between January 2022 and August 2026, taking windows from 30 minutes before each release to 60 minutes after it. Release-free days supply same-clock-time placebo windows. The shared likelihood covers 732 windows (430 release, 302 placebo) and 7.3 million scored events.

Two clocks emerge. Among the five release types that pass the placebo test, the fitted direct news kernel reaches half its response in a median 1.1 s. The pooled fit's branching-matrix spectral radius is 0.87, below the stability bound of one. A Markov embedding adds the trading set off by that initial reaction in closed form: half the extra activity arrives within 16 to 137 s, and 90% within 5 to 11 minutes. The paper also develops EPT-X, a GRU-driven neural model on the same Erlang state space with a mixture-of-Erlang renewal channel for event timing. Its likelihood remains exact.

Can MSX be rebuilt?

The estimator is specified closely enough to rebuild. Endogenous kernels have 10 rates spanning 5ms to 600s at Erlang order 2. News kernels have 8 rates from 0.1 to 300s at order 3; their marks are a constant and the positive and negative parts of standardised surprise. Anticipation kernels use 4 rates from 10 to 300s. Each window gets a linear baseline trend, alongside 96 time-of-day hat functions. On synthetic data, the certificate stayed at or below 3.3x10^-4 nats across 100 fits. An implementer can reconstruct the optimizer from the description, and the author releases code for comparison.

The choices inside the event clock

Delta comes first. It is chosen separately for each asset to produce a median rate of about 0.2 events per second. Absorption times consequently refer to crossings of a threshold tuned to that rate. The surprise measure also requires a choice. For payrolls, claims, PPI and retail sales, Dutta says the expectation is "a statistical forecast from earlier first prints" rather than survey consensus; we did not find the forecasting model spelled out. Placebo days exclude "sixteen second-tier US releases", and we did not find that list either.

The l1 penalty on news kernels has the largest effect. Held-out likelihood on the 2024 windows selects a peak at 0.8, where mean release mass is 1.81 versus placebo mass of 1.03. At 0.2, those masses are 10.79 and 6.23. The author reports two earlier specifications that failed the placebo test: one absorbed 169 extra events per pseudo-release, the next 18.6. The author says the final specification "was chosen because it passed this test, not because of the results it produced." Disclosing that history is good practice. The final result still follows a three-step search guided by the placebo test, with the last step also changing which days qualified as placebos.

What passes the placebo test?

The direct news kernel mostly passes. Five of eleven release types clear the 50% threshold: CPI at 8.5x, payrolls 6.7x, PPI 4.0x, the FOMC statement 3.9x and retail sales 1.9x. GDP, claims, ECB and BoJ decisions and the press conferences fail. We did not find standard errors for these ratios. Retail's 1.9x clears the 50% bar narrowly enough that a resample could erase the margin.

Echo-corrected time measures something else. Its t90 of 5 to 11 minutes falls just short of the slowest echo mode, which decays in about 12 minutes. The paper prints echo-corrected times for the release types that fail the placebo test, too. They occupy much the same band. The long tail therefore mostly tracks market self-excitation after a shock of any kind, consistent with the author's conclusion that absorption speed "is not a property of the news kernel alone." The fraction Π(t) sums up and down events. Two-way crossings can fill that measure without moving price anywhere.

The certificate's reach

The pooled fit has a gap of 0.0054 nats per dimension. It bounds optimizer error for the chosen specification; it cannot establish that the specification describes the market. Per-window excess-dispersion tests still reject in 26 to 55% of windows. The paper sets this beside Rambaldi et al.'s 42% and 93% of trading days, though trading days are a different unit. At 0.87, the spectral radius is near-critical. The paper cites work showing that mis-specified intraday seasonality can bias branching ratios upward. Its 96 time-of-day hat functions and per-window trend do not remove that concern; after the first specification, seasonality is modelled in New York time. The yearly readings are 0.91, 0.89, 0.85, 0.86 and 0.85 for 2022 to 2026, flat after 2024. FX data also have a gap in the first half of 2023. Those qualifications matter before treating the series as falling endogeneity.

EPT-X wins on likelihood

On LOBSTER, EPT-X records 3.534 nats per event for AAPL against 3.370 for nonparametric Hawkes EM. For INTC it records 6.163 against 5.713 for IntensityFree. The 5.713 is the corrected figure: EasyTPP ships IntensityFree with a clamp on gaps below 10^-5 s, producing a reported 6.790 and affecting 29% of INTC test gaps. Catching that clamp helps anyone comparing point-process models. Even linear MSX-auto reaches 3.361 on AAPL, versus 2.315 for S2P2.

The sample is one trading day, 21 June 2012, as the author acknowledges. Tied timestamps were separated by a uniform offset in (0, 100ns), leaving every model to score artificial sub-microsecond gaps.

The evidence closest to trading is less persuasive. Trained on 2022 to 2024 and tested on 163 releases in 2025 and 2026, MSX forecasts of log(1+N) post-release crossings have RMSE of 1.322, 1.049 and 0.955 at 1, 5 and 15 minutes. MSX improves on climatology (1.440, 1.308, 1.199). It trails event-study OLS (1.140, 0.925, 0.776) and HAR (1.217, 0.985, 0.820) at every horizon. The paper argues that those regressions are fitted directly to the target. The structural model still loses on activity. Its one-minute drift forecast correlates 0.17 with realised drift, compared with 0.06 for the best baseline. We did not find a P&L, a cost assumption or a position rule anywhere.

We could not test this ourselves. Our finest data are one-minute bars; delta-crossing events require millisecond quote or trade timestamps. The application also calls for event-level expectations for each release, which our macro series lack.

A 0.17 drift correlation across 163 test releases would make this a trading paper if it survived a one-minute entry and realistic spreads.