ai-asset-pricing: Empirical Asset Pricing Tools (Dickerson)

  • Author/Source: Alexander M. Dickerson, GitHub repo

  • Original: https://github.com/Alexander-M-Dickerson/ai-asset-pricing

  • Key Ideas

  • A multi-agent-friendly empirical asset pricing research repo: clone it, ask Claude Code (/onboard), Codex, or Gemini CLI to onboard, and you reach a research-ready WRDS workflow without hand-editing machine-specific config inside the shared tree.
  • Three parallel adapters cover the major coding agents: CLAUDE.md (Claude), AGENTS.md (Codex), GEMINI.md (Gemini CLI), all routing through a shared tools/bootstrap.py audit/apply engine. The shared cold-start scripts are tools/onboard.ps1 (PowerShell) and tools/onboard.sh (bash).
  • Coverage spans the full empirical-asset-pricing stack: WRDS data extraction, factor-model setup and reusable automation, a fintools/ shared utility package, the PyBondLab portfolio-construction package, and LaTeX paper scaffolding/auditing/compilation.
  • Local state is kept outside the repo in a per-user OS-specific directory (canonical files: local_env.md, claude.local.md, settings.local.json). This is explicitly designed to make the repo safe in a shared Dropbox/OneDrive working tree --- machine-local paths don't sync across users. Repo-root compatibility shims (CLAUDE.local.md, .claude/settings.local.json) are legacy and only generated on request.
  • WRDS access is optional. The agent asks once at onboarding; without WRDS, the repo still supports skills/rules/agents, fintools, PyBondLab install with bundled-data smoke tests, and LaTeX boilerplate.
  • Manual WRDS bootstrap files are documented: ~/.pg_service.conf, ~/.pgpass (with chmod 600), and ~/.ssh/config Host wrds entry for SSH/TAQ workflows.

  • Acknowledges Piotr Orlowski's claude-wrds-public agents and Scott Cunningham's MixtapeTools as starting points, both heavily augmented.

  • Summary

Dickerson's ai-asset-pricing repo is best understood as a template for setting up a Claude-Code-friendly empirical asset pricing project, with parallel support for Codex and Gemini CLI. The design choices reflect lessons that show up across this wiki: agent context lives in markdown (CLAUDE.md, AGENTS.md, GEMINI.md), reusable workflow content lives in plain markdown under .claude/agents/, .claude/skills/, .claude/rules/ so it can be reused across tools, and a single shared bootstrap.py engine handles the actual Python-environment setup so the per-tool adapters stay thin.

The most consequential design decision for finance researchers working in shared file systems (Dropbox/OneDrive) is that canonical local state is held outside the repo. This avoids the well-known footgun where a coauthor opens the repo on a different machine and Claude reads stale paths from a synced .claude/settings.local.json.

The functional surface includes WRDS-oriented data work, factor-model research scaffolding, PyBondLab for portfolio construction, and LaTeX writing infrastructure. Onboarding is tools/bootstrap.py audit → execute the emitted bootstrap_plan → apply → re-audit, with cross-platform entry points (onboard.ps1 / onboard.sh).

  • Relevance to Economics Research

Direct and high. This is a working reference implementation of "what a Claude-Code-ready empirical finance project looks like" --- the AssayingAnomalies/finance-research analog. Anything I (or anyone running an empirical asset-pricing project on WRDS) would build looks structurally similar to this. Specific patterns worth borrowing: external local-state files for Dropbox safety; parallel adapters for Claude/Codex/Gemini so the project isn't locked to one vendor; bootstrap.py audit/apply as the shared environment-setup engine; the conditional WRDS prompt at onboarding so the repo also works for collaborators without WRDS.