HPIN barely moves a scalar adverse-selection control. Diniz, Curtis, Martins and Sbruzzi add a hidden Markov chain to PIN, yet the authors conclude that the estimated level changes by only a few thousandths. In their overdispersed simulations, both estimators overstate the true PIN of 0.1163. Mean absolute error is 0.0157 for HPIN and 0.0193 for classical PIN. HPIN comes closer in all six environments, though the difference is too small to change a regression control. Its real addition is an estimate of how long an informed-trading regime lasts.

Where duration enters

Classical PIN draws each day independently. Nothing happens with probability 1 minus α; otherwise the news is good (weight 1 minus δ) or bad (weight δ). Uninformed buys and sells follow Poisson arrivals at rates ϵb and ϵs. Informed traders add μ to the side matching the news, giving PIN as αμ/(αμ + ϵb + ϵs). Independent days leave duration out of the model.

The authors' HPIN (Hidden Probability of Informed Trading) retains those three states and lets a Markov chain carry them across days. Its transition matrix uses six economic parameters: the news-arrival hazard η, the bad-news share κ, persistence for good and bad regimes (φ+, φ−), and the chances that a signal flips sign (ω+, ω−). The likelihood ties each state to its order-flow rates. Neutral has (ϵb, ϵs); good news adds μ to buys, and bad news adds μ to sells. Re-imposing those ties at every M-step keeps the labels fixed, without the post-hoc clustering Yin and Zhao needed for their generic states. Static PIN is the case in which every transition-matrix row equals the stationary vector. The authors distinguish that case from the identity matrix, which would leave each regime permanent.

The quantity to watch is the spectral gap, one minus the second-largest eigenvalue modulus of the transition matrix. A gap of 1 means the market forgets its state overnight. A small gap means regimes persist.

Every result comes from synthetic data. Each environment consists of one series of T = 5,000 periods, with ϵb = ϵs = 10,000, μ = 5,000 and seed 42. Across six environments, persistence ρ runs from 0 to 0.9. The true gap is 1 minus ρ, while unconditional PIN remains 0.1163. Each environment has a pure Poisson run and a run with a common Inverse Gaussian multiplier (mean 1, scale 9) on both sides. That multiplier overdisperses the counts, leaving both models with misspecified Poisson emissions.

Enough detail to rebuild the estimator

The appendix specifies closed-form quadratic roots for ϵb, ϵs and μ. Each root is conditional on the other parameters, and the procedure cycles through them once per iteration. It uses a pseudocount of 10^-3, computes emissions in log space, and sets tolerance at 10^-8 with a 1,000-iteration cap. Fits converge in under twenty iterations. The cold-start transition matrix assigns 0.8 to neutral-to-neutral; every rolling window (60 to 180 periods) starts there again. The static comparison uses PINstimation with the Ersan and Alıcı starting grid, a fair benchmark.

The worked example supplies a checksum. At w = 180, observation 1,000, its predictive log score difference is −0.147226, enough to check a reimplementation to six decimals. Across 48,800 production rolling fits, none reached the μ = 0 boundary where the states collapse.

Does the gap survive overdispersion?

With correct Poisson emissions and w = 180, the gap has mean absolute error of 0.037 and correlation to truth of 0.999. PIN comes within 0.0022. Static PIN returns 1.000 in every environment, for an error of 0.492. The authors call that constant "an arithmetic consequence of its degenerate transition matrix," and the useful comparison for the gap is therefore HPIN against the truth.

Overdispersion flattens the estimate. Error climbs to 0.255, and the estimated gap spans only 0.34 of the true 0.90 range, although its correlation with truth remains 0.992. The authors describe the setting as "a realistic stress test," and acknowledge that the gap works as a ranking rather than a cardinal measure. They also note that, even compressed, it is roughly twice as close to truth as the static constant (0.255 against 0.492). The scale remains a problem: a sticky informed regime with a true gap of 0.100 reads 0.560, suggesting moderate switching.

Shorter windows make that problem worse. Even with correct emissions, recovered amplitude falls from 0.817 at w = 180 to 0.690 at w = 60. For memoryless data with overdispersed emissions, the rolling gap is 0.902 at w = 180 and 0.827 at w = 60. We did not find repetitions across seeds, so these errors have no reported sampling band.

There is also a reading hazard in the parameter results. The prose says HPIN systematically gives higher ϵb, while Table 10 puts HPIN's mean ϵb lower at every window: 9,916.25 against 9,934.78 at w = 60. The paper's prose disagrees with itself on dispersion too.

Forecasts stay level

The one-step-ahead, out-of-sample predictive log score is a draw. At every window, the median paired difference stays within ±0.008. Wilcoxon rejects only at w = 60 (p = 0.023); from 90 through 180, the p-values are 0.57, 0.12, 0.29 and 0.16. HPIN wins on 48.7% to 50.6% of days.

It wins on 46.6% of days when the true regime switched, compared with 51.9% on other days.

The authors concede the forecasting result and rest their case on HPIN's ability to represent regime duration. That argument holds under clean Poisson data. In the overdispersed data they call realistic, the extra quantity survives as a ranking, compressed into a gap band of roughly 0.56 to 0.90, as they acknowledge.

Before a desk can use it

We could not run HPIN ourselves. Estimation needs daily counts of buyer-initiated and seller-initiated trades; our data consists of OHLCV bars, without trade prints or trade sign. The paper lists trade-direction classification among the steps required for real data and leaves it to future work. An implementer would also have to contend with classification error.

Our own speculation, untested in the paper, is that a desk could use the gap as an execution-risk flag: a small estimated gap suggests an informed regime tends to persist. The paper tests no strategy and reports no returns. It also leaves validation against spreads, volatility and price impact to future work. A real-data run in which the ordinal gap, estimated on 60-day windows, leads spread widening would change my view.