Belief Aggregation¶
Pool several probability assignments over a common event space into one collective distribution under an explicit rule.
Core Idea¶
Belief aggregation in its probabilistic sense turns several probability assignments about the same events into one collective probability assignment through a specified pooling rule. Inputs may be expert judgments or probability forecasts from statistical and institutional sources. With fixed nonnegative weights summing to one, a linear pool averages the source probabilities for each event; a geometric or logarithmic pool combines positive probabilities across possible worlds multiplicatively and renormalizes. These are different rules, not one inevitable consensus formula.[^ref-abf2af3c3f0e]
The output is a probability distribution, not a majority vote or ranked list. Pooling does not by itself establish truth, calibration or collective wisdom; the question, weighting rationale and evaluation criterion must be supplied separately.[ref-abf2af3c3f0e][ref-b72d95e8a9cd]
Scope of Application¶
The sources must address one common event agenda or outcome variable under comparable definitions and horizons. Sangay researchers elicited expert uncertainty distributions on shared volcanic target questions and compared equal and performance-based weights to form group decision-maker distributions. Ranjan and Gneiting combined statistical and National Weather Service probabilities for precipitation using a beta-transformed linear pool. The clinical-guideline voting example in the frozen seed is excluded because it did not pool treatment-effect distributions.[ref-49b33166ff1a][ref-b72d95e8a9cd]
“Probabilistic opinion pooling” is an attested scoped alternate name. Broad “opinion aggregation” and “risk aggregation” may refer to voting, ranks or portfolio losses and remain unresolved as aliases. Geometric-pool positivity assumptions and any forecast-calibration claims must be checked for the chosen implementation.[^ref-abf2af3c3f0e]
Clarity¶
The identity fixes the carrier: probability reports about the same events. Rank aggregation combines ordinal orders; belief revision changes one agent's state with new information; Dempster–Shafer theory combines a different subset-mass representation. An ensemble-learning system may supply forecasts, but training multiple predictors is not a necessary part of pooling those forecasts.[^ref-abf2af3c3f0e]
The proposed DAG parent is live Aggregation: several probability reports are deliberately reduced to one. Live Wisdom of the Crowds is not a strict parent, because independent errors and improved accuracy are not guaranteed by a pool.
Manages Complexity¶
One pooled distribution can stand in for many source distributions during a forecast or decision analysis. This simplifies the downstream interface but can hide between-source disagreement and the effects of alternative weights; retain the input reports when those differences matter. The Sangay paper compared pooling schemes, while the weather paper studied recalibration precisely because arithmetic combination can alter forecast performance.[ref-49b33166ff1a][ref-b72d95e8a9cd]
Abstract Reasoning¶
Align events and time horizons; verify each input is a coherent probability assignment; state the pooling rule and weights; compute and normalize the output; then test the chosen objective. Dietrich and List show a real rule trade-off: linear pooling satisfies eventwise independence with unanimity in their classical domain, while geometric pooling can commute with a common Bayesian update under its regularity conditions but is not generally eventwise independent. A rule's mathematical properties and a pooled forecast's empirical calibration are different tests.[^ref-abf2af3c3f0e]
Knowledge Transfer¶
The Sangay and precipitation cases share the same roles: common target, multiple probability reports, explicit pooling rule, and one collective assignment. Sangay's calibration questions can support expert weights; precipitation forecasting uses predictive performance and recalibration. Those setting-specific checks do not transfer automatically. The portable many-to-one operation belongs to live Aggregation, while common-event probability coherence keeps this named abstraction domain-specific.[ref-49b33166ff1a][ref-b72d95e8a9cd]
[^ref-abf2af3c3f0e]: Franz Dietrich and Christian List, “Probabilistic Opinion Pooling”, original author manuscript, final 13 October 2014, introduction and §§2, 4–7, PDF pp.1–14. [^ref-49b33166ff1a]: Benjamin Bernard et al., “Developing hazard scenarios from monitoring data, historical chronicles, and expert elicitation: a case study of Sangay volcano, Ecuador”, Bulletin of Volcanology 86, article 68 (2024), “Expert elicitation basics” and “Elicitation results.” [^ref-b72d95e8a9cd]: Roopesh Ranjan and Tilmann Gneiting, “Combining Probability Forecasts”, Journal of the Royal Statistical Society: Series B 72(1) (2010), 71–91, publisher abstract and indexed introduction; full text was not directly readable during this author pass.
Relationships to Other Abstractions¶
Current abstraction Belief Aggregation Domain-specific
Parents (1) — more general patterns this builds on
-
Belief Aggregation is a kind of Aggregation Prime
Pooling probability assignments is a deliberate many-to-one aggregation with probabilistic coherence constraints.
Hierarchy path (1) — routes to 1 parentless root
- Belief Aggregation → Aggregation → Micro Macro Linkage
Neighborhood in Abstraction Space¶
Belief Aggregation sits in a sparse region of the domain-specific corpus (81st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Strategic Decision Biases & Mechanisms (29 abstractions)
Nearest neighbors
- Scoring Rule — 0.83
- Boltzmann Fair Division — 0.83
- Shared Information Bias — 0.82
- Lottery (decision theory) — 0.82
- Factored Language Model — 0.81
Computed from structural-signature embeddings · 2026-10-08