Skip to content

Latest commit

 

History

History
16 lines (13 loc) · 892 Bytes

File metadata and controls

16 lines (13 loc) · 892 Bytes

Forecast Bench v6

LLM-based agentic forecasting system for binary questions. Supports backtesting on AIBQ2 (Metaculus) and ForecastBench (markets + datasets), plus live competition submission.

Context

  • docs/workflow.md — full pipeline walkthrough
  • docs/TODO.md — tracked TODOs for paper and code
  • src/MODULES.md — categorized guide to all source modules
  • docs/shrinkage.tex — logit-space shrinkage derivation
  • ../nips26/main.tex — NeurIPS 2026 paper draft (moved to top-level)

Layout

  • src/ — all Python source (see src/MODULES.md for module guide)
  • data/ — inputs: questions, exams, tags, caches
  • experiments/ — xid definitions + outputs (forecasts, eval)
  • Configs use hyphenated directory names (e.g. flash-high-brave-c1-t1); delta strings (e.g. flash/thk:high/crowd:1) are used on the CLI but resolved to directory names internally