ralph-wiggum-asset-pricing: Open-Source Loop for Generating Asset Pricing Papers

  • Author/Source: Andrew Y. Chen (Federal Reserve Board), GitHub repo

  • Original: https://github.com/chenandrewy/ralph-wiggum-asset-pricing

  • Key Ideas

  • Open-source framework that generalizes the workflow used to generate "Hedging the Singularity" (see summaries/prompts-to-paper). Implements Geoff Huntley's "Ralph Wiggum loop": plan → improve → test → repeat, with a human staying on the loop rather than in it.
  • Each iteration: ralph/author-plan.py reads test results + spec/paper-spec.md and writes ralph-garage/improvement-plan.md; ralph/author-improve.py executes the plan; tests/*.py evaluates; failures send the loop back to step 1.
  • Branch model: humans work on main, Ralph works only on ralph/run, with each iteration being a single rloop-NN: commit. This makes stopping, resuming, and discarding stretches cheap. go-ralph-go.sh from main starts/extends a stretch; from ralph/run it resumes.
  • YOLO mode: Ralph invokes Claude with --dangerously-skip-permissions and Codex with --sandbox danger-full-access. The recommended sandbox is the included dev container (Docker), which isolates the YOLO agents and provides bash, git, Python, R, LaTeX (pdflatex, biber), Poppler, and the agent CLIs.
  • Multi-agent support via three module-level constants per script (AGENT, MODEL, EFFORT); checked-in defaults are Claude. Each test/referee can use its own agent/model/effort tier.
  • Test taxonomy (the full 25-test version on ralph/run-final for "Hedging the Singularity"): element-* (required content present), factcheck-* (claims match code/lit/each other), spec-* (paper matches spec), theory-* (clarity, no deadweight, intro pay-off), visual-* (figures render), writing-* (intro and intuition prose). Plus a build-latex infrastructure test.
  • Ships with a blank template; users replace spec/paper-spec.md and spec/economic-background.md with their own paper. Optionally bash check-ralph-direction.sh to have AIs generate candidate starting points first.
  • tests/ for the asset-pricing theory paper is a small subset; for empirical papers Chen points to HumanxAI-ChenAY's test set, with families like factcheck-econ, story-narrative, story-exhibit-coherence, transparency-calibration, visual-tables. Treat them as source material to port, not files to copy.
  • Cost reference point: "Hedging the Singularity" with all 25 tests burned through two $200/month Claude Code subscriptions end-to-end. The default main test set is much lighter.

  • WRDS credentials handled via host-side python .credentials/setup.py, exposed inside the dev container as WRDS_USERNAME / WRDS_PASSWORD environment variables.

  • Summary

This repo is the productionized, generalizable version of the workflow Chen used in "Hedging the Singularity." Where Prompts-to-Paper is a frozen archive of the prompts that produced one paper, ralph-wiggum-asset-pricing is a reusable scaffold: clone it, replace the spec and tests, hit bash go-ralph-go.sh, and the agents will iterate on a paper draft until your tests pass.

The interesting design choices are about controlling agent behavior at scale. The two-branch model (main for humans, ralph/run for the agent) treats agent output as cheap and disposable: a stretch that goes badly is a git branch -D ralph/run away from being undone. Tests are first-class: failures drive the loop, so the test suite is essentially the spec for the paper's quality bar. The honesty about cost (two $200/month subscriptions for one paper) is unusual and useful as a planning anchor.

Chen's preface to "Hedging the Singularity" (linked from the README) reports that an attempted "human as Clockmaker" mode --- setting everything up correctly upfront and then leaving Ralph alone --- did not work. The lesson seems to be that paper-quality output still requires active human steering of the spec, the test mix, and the prose direction. This complements (and tempers) the headline claim from prompts-to-paper.

  • Relevance to Economics Research

Highly relevant for finance/economics researchers who want to experiment with paper-generation pipelines. Three things make this concrete in a way most "AI for research" discussions aren't: (1) full source for the loop, the prompts, and the test suite is published; (2) the recommended cost ($200/mo × 2 subscriptions) is a real number to plan against; (3) the test taxonomy is reusable scaffolding even if you never run the full loop --- adopting just the factcheck-code, factcheck-lit, and spec-paper tests as a CI step on a human-written paper is a low-cost experiment.

For asset-pricing/empirical-finance work specifically, the pointer to HumanxAI-ChenAY's test families (factcheck-econ, story-narrative, transparency-calibration, visual-tables) is more relevant than the bundled theory tests --- those are the failure modes empirical finance papers actually have.