Iran Supreme Leader Succession — Bayesian Expert Aggregation (v2)

Fourteen Iran specialists. Seventy-six scored predictions stretching back to 2006. One hundred thousand Monte Carlo draws. A precise mapping from each expert's verbatim statement on succession to a probability vector over ten candidate outcomes — and an accuracy-weighted, correlation-corrected consensus on who follows Khamenei.

On February 28, 2026, Ayatollah Ali Khamenei was killed in joint US–Israeli airstrikes. Three days later, under Article 111 of the Iranian constitution, an Interim Leadership Council took the reins of the Islamic Republic. The question every Iran watcher is now asking — who becomes the next Supreme Leader? — has no clean answer. Fourteen expert analyses give fourteen overlapping but distinct pictures, and a simple average of their views throws away the thing that matters most: some of these experts have been right about Iran for two decades, others have been conspicuously wrong.

This report builds a Bayesian aggregation over the panel. Each expert is weighted by their demonstrated accuracy on a scored history of Iran predictions. Each expert's most recent verbatim statement on succession is read carefully, mapped onto ten candidate outcomes via explicit ruled-out / ruled-in reasoning, and converted into a Dirichlet probability vector. The vectors are aggregated with an accuracy-weighted linear pool, then extremized to correct for shared information across the DC think-tank and IRGC-specialist clusters. The result is a posterior distribution with 80% and 95% credibility intervals over every named successor.

Two security-credentialed clerics — Mohseni-Ejei and Arafi — lead the field, together absorbing roughly 39% of the posterior mass. The highly-discussed heir-apparent Mojtaba Khamenei is only fourth, at 12.3%: the highest-accuracy experts systematically discount him as either too independently powerful for the IRGC to control, or too weak theologically to carry the mantle. The largest tail-risk outcome is not any individual but the Leadership Council scenario — an IRGC-dominated collective or de-facto abolition of the office — at 14.0% with the widest posterior of any category.

How the model works

The aggregation rests on four ingredients: a historical accuracy score for each expert, a precise reading of what each expert has actually said about succession, a Dirichlet distribution capturing their uncertainty, and a correlation-aware Monte Carlo pool.

Step 1 — expert reliability as a Beta posterior

Each expert's track record is scored on a five-point scale: TRUE (1.0), MOSTLY TRUE (0.75), PARTIAL (0.5), MOSTLY FALSE (0.25), FALSE (0.0). These feed a Beta posterior built on a Jeffreys prior:

α = (Σ scores) + 0.5 ; β = (N − Σ scores) + 0.5 ; E[p] = α / (α + β)

The posterior mean E[p] is each expert's expected accuracy. The Beta posterior preserves uncertainty: an expert with many predictions has a narrow posterior; one with few predictions, a wide one. A newcomer with three correct calls is not trusted like a veteran with twenty.

Step 2 — prediction-to-candidate mapping

This is the step that distinguishes v2 from the earlier report. Each expert's most recent verbatim statement on succession is analyzed to produce three things: (a) which candidates the statement explicitly or implicitly rules out, with reasoning traced directly to the text; (b) which candidates are ruled in as plausible; and (c) a probability weight assignment. When the expert expresses no preference among plausible candidates, they receive equal weight. When language implies relative likelihood — "sole frontrunner", "most likely", "mediocrity rewarded" — weights are adjusted accordingly, and the inferential chain is documented.

Step 3 — Dirichlet concentration

Each expert's weight vector is modelled as a Dirichlet distribution with concentration parameter κ. Higher κ (12–16) is assigned to experts who name specific candidates with confidence; lower κ (6–10) to structural or generic predictions. The Dirichlet preserves uncertainty around the point estimate when Monte Carlo sampling the consensus.

Step 4 — Monte Carlo aggregation with correlation discount

For each of 100,000 draws the model (1) samples expert reliability from the Beta posterior, (2) samples candidate weights from the Dirichlet, and (3) computes an accuracy-weighted linear pool. A 14×14 correlation matrix discounts shared information — DC think-tank cluster at 0.35–0.50, IRGC specialists at 0.30–0.45, hawkish analysts at 0.35–0.45. Eigenvalue decomposition yields 10.4 effective independent experts out of 14 and an extremizing factor d = 0.69 (Satopää et al. 2014).

The expert panel

Fourteen specialists, ranked by posterior mean accuracy. N is the number of scored historical predictions in the tracker; Σ is the sum of scores; α and β parameterise the Beta posterior.

The panel spans every major analytical camp — Carnegie-Brookings-CFR (Sadjadpour, Maloney, Vaez, Slavin), the IRGC specialists (Ostovar, Alfoneh, Aarabi), the diaspora intellectual network (Nasr, Khalaji, Milani), and the hawkish camp (Takeyh, Rubin). Within-camp correlations are 0.30–0.50; cross-camp correlations 0.05–0.20.

What each expert actually said

The core of the v2 methodology is the verbatim-to-weights mapping. Below, for each of the fourteen experts: the quote, which candidates their statement rules out and why, which are ruled in, and the resulting weights. This is where the analytical work is — the rest is arithmetic.

The consensus forecast

The consensus is computed as the accuracy-weighted linear pool of the 14 expert vectors, each weighted by its Beta posterior mean E[p]. The correlation adjustment then extremizes the result using factor d = 0.69 (Satopää et al. 2014), correcting for shared information across the four analytical camps. Highest Density Intervals are computed from the Monte Carlo posterior across 100,000 draws.

Why these numbers

Mohseni-Ejei at 20.3% is the panel's plurality pick. Six experts give him ≥20% weight: Sadjadpour (20%), Ostovar (25%), Khalaji (35%), Nasr (35%), Takeyh (30%), Slavin (20%). He fits two distinct archetypes — Khalaji's "weapon + prison key" (Chief Justice, ex-Intelligence Minister) and Nasr's "hardline cleric after assassination." Only Aarabi, Rubin, and Vaez give him as low as 10–12%.

Arafi at 18.5% draws his strongest support from Vatanka (30%), Alfoneh (30%), Sadjadpour (25%), Ostovar (25%), Khalaji (22%). He is the canonical "obscurity promoted" candidate (Sadjadpour), the "artless" controllable cleric (Alfoneh), the bureaucrat with Khamenei's confidence (Vatanka). Nasr, Rubin, and Vaez give him less weight — they are looking for either a hard-line cleric or a structural break, and Arafi is neither.

Leadership Council at 14.0% is driven almost entirely by Vaez (50%) and Milani (22%), with Rubin's 22% for "military junta" as a third contributor. Vaez's "last supreme leader" prediction, anchored at 84.4% accuracy, is the single statement doing most of the work here. But ten of fourteen experts give Council ≤12%, which keeps it from dominating — while also making it the highest-variance outcome in the whole forecast (95% HDI runs from 0.2% to 37.9%).

Mojtaba at 12.3% is the most important story in the model. Aarabi gives him 50% — the single strongest candidate endorsement in the panel — and Slavin gives 25%, Rubin 18%. But the seven highest-accuracy experts all independently discount him: Sadjadpour (trigger unrest), Khalaji (too powerful for IRGC to control), Alfoneh (not artless), Nasr (hardline cleric, not mid-level), Vaez (3%), Ostovar (10%), Milani (10%). The conventional-wisdom frontrunner is systematically downweighted by everyone except his primary advocates.

H. Khomeini at 5.8% — only Takeyh (22%) and Maloney (15%) give him significant weight, both citing his appearance on Khamenei's nominee list. Nasr's assassination-conditional rules directly against him, Khalaji rules him out on credentials, and eight experts give him ≤5%.

Mirbagheri at 7.3% picks up weight primarily from Nasr (22%) via the "assassination → ultra-hardliner" conditional, with meaningful support from Sadjadpour, Khalaji, and Ostovar. He fits the hardline-cleric archetype but his public profile beyond the Paydari Front is thin.

Bushehri at 7.6% accumulates across multiple experts as the second-tier "weak, controllable cleric": Alfoneh (15%), Ostovar (13%), Khalaji (12%), Maloney (12%). No expert gives him the lead, but he absorbs consistent secondary weight.

The remaining mass — Larijani at 5.2% (kingmaker/broker in Vaez and Wright's frameworks), Placeholder at 4.0% (almost entirely Rubin's Brezhnev–Andropov parallel, from the panel's lowest-accuracy expert), and Other / Unknown at 4.9% — captures residual tail risk across scenarios no single expert emphasises.

It is quite possible that Ayatollah Khamenei is the last supreme leader of Iran. Ali Vaez · International Crisis Group · NPR, July 2025

The single most provocative statement in the panel belongs to Vaez, and it is the reason Leadership Council sits at 14% rather than at 4% where the seven other structural-pessimists collectively place it. Vaez is not a contrarian outlier — he is an 84% accuracy forecaster whose 2025 call on US–Israeli strikes already resolved TRUE. When a top-tier expert says the office itself may be ending, the posterior has to take it seriously.

Model parameters