Skip to content

Agentic Forecasting using Sequential Bayesian Updating of Linguistic Beliefs

Home / Papers / Agentic Forecasting using Sequential Bayesian Updating of Linguistic Beliefs

Agentic Forecasting using Sequential Bayesian Updating of Linguistic Beliefs

Authors: Kevin Murphy Date: 2026-04-20 Paper ID: openalex:2604.18576

Summary

BLF (Bayesian Linguistic Forecaster) is an agentic framework for binary forecasting that improves decision-making by replacing monolithic context window accumulation with a structured Bayesian belief state. This state effectively fuses numerical probability estimates with iterative natural-language evidence summaries generated via tool-use. Furthermore, the system incorporates hierarchical shrinkage aggregation and calibration techniques to handle uncertainty and base-rate skew. Evaluations on 400 backtesting questions from the ForecastBench benchmark demonstrate that BLF outperforms leading models while maintaining a rigorous, low-leakage testing protocol.

Key Contributions

  • Introduced the BLF agentic forecasting system, achieving state-of-the-art performance on the ForecastBench benchmark.
  • Developed the Bayesian linguistic belief state, a memory-efficient semi-structured representation for LLM iterative reasoning.
  • Proposed hierarchical multi-trial logit-space shrinkage and hierarchical Platt scaling for robust probability calibration.

Open Questions & Future Work

Key Concepts

  • bayesian-linguistic-belief-state: A semi-structured state representation that combines numerical probabilities with natural-language evidence summaries, maintained through iterative updates.

Archivist Review

The ‘Bayesian Linguistic Belief State’ is approved as a significant contribution to agentic memory management in long-horizon forecasting. I renamed the open question to be more descriptive of the scientific uncertainty regarding architectural transferability and rejected the tool-specific question as it reflects common implementation refinement rather than a core research bottleneck. The ForecastBench dataset is approved as a primary evaluation tool for this class of model.

Approved Concepts

  • Bayesian Linguistic Belief State: It introduces a novel way to manage agentic memory and reasoning by explicitly decoupling belief representation from raw context window growth, a critical bottleneck in long-horizon forecasting.

Approved Open Questions

  • Agentic Forecasting Model Transferability: Understanding model-agnosticism is critical for determining the scalability and robustness of the agentic forecasting framework across different computational environments.

Rejected Candidates

  • [open_question] Enhancing Domain-Specific Forecasting Tools (domain-specific-tool-development-forecasting) - low_impact: This is essentially a request for better tools/data features rather than a fundamental scientific research question about the forecasting mechanism itself.

Datasets

Metadata & Links

url
https://arxiv.org/abs/2604.18576
paper_id
2604.18576
paper_source
openalex
domain
nlp
tags
forecastingllm-agentsbayesian-inferencecalibration
architectures
datasets
forecastbench
concept_slugs
bayesian-linguistic-belief-state
dataset_slugs
forecastbench
skill
TimeSeriesSkill
created_at
2026-04-23T05:47:33Z