HUNTER×HUNTER chapter 421

Two exploratory models: historical interruptions and a scenario informed by Togashi’s manuscript reports. Probabilities depend on their assumptions.

Reference: historical interruptions

Model median

June 2027

Central 80% predictive interval

November 2026 – March 2031

Calculated Oct 27, 2026, 9 interruptions, 3 with exact calendar dates.

80% probability within a model does not establish 80% real-world reliability. Data are sparse and editorial decisions remain unknown.

Modelled probability of return within the next…

90 days

27%

180 days

43.2%

365 days

61.9%

730 days

77.9%

Reference: historical interruptions

Historical reference: every completed interruption since 2014 between consecutive numbered chapters, including short breaks. Duration runs from the last chapter to the return. Exact dates take precedence; otherwise assume 48 Jump issues per year.

Log duration is normally distributed. A normal–inverse-gamma prior represents uncertainty in its mean and variance. Integrating both yields a Student-t predictive distribution, conditioned on no return by the calculation date. Recomputed at each site build.

yᵢ = log(Tᵢ) ~ Normal(μ, σ²)
σ² ~ InvGamma(α₀, β₀) ; μ | σ² ~ Normal(μ₀, σ²/κ₀)
κₙ = κ₀+n ; μₙ = (κ₀μ₀+nȳ)/κₙ
αₙ = α₀+n/2
βₙ = β₀ + Σ(yᵢ−ȳ)²/2 + κ₀n(ȳ−μ₀)²/(2κₙ)
log(Tnew) | data ~ t₂αₙ(μₙ, √[βₙ(κₙ+1)/(αₙκₙ)])
P(T ≤ t | T > e, data) = 1 − S(t)/S(e)

μₙ = 5.3454 ; κₙ = 9.25
αₙ = 6.5 ; βₙ = 11.7051
df = 13 ; scale = 1.4126

Parameters, priors and sensitivity

The historical prior centers log(T) on log(365.2425 days), with κ₀ = 0.25, α₀ = 2 and β₀ = 1. Centering on 180 or 730 days instead gives medians of June 2027 and June 2027.

Using eligible history since 1998 gives a median of December 2026.

Short and long interruptions are heterogeneous; one distribution cannot describe every regime change. The 2014 cutoff and issue-to-day conversion are assumptions, not learned parameters. No medical information is inferred from posts. An official date hides these estimates.

Chronological validation on 6 returns: mean absolute error 436 days; 5/6 inside the 80% interval. Each prediction uses only previously completed interruptions.

Comparison with a fitted log-normal without parameter uncertainty: error 442 days. Mean log score: Bayes -7.9, reference -7.92 (higher is better). So few returns cannot establish superiority.

Sources and complete hiatus history