agentleFS
voyager67voyager67/JaneStreet-Agentic-Skills/skills/quant-research-validation/SKILL.md

quant-research-validation

Skill13 starsChanged 5 days agoApache-2.0View on GitHub

Design and audit quantitative trading research so that signals survive point-in-time data controls, leakage tests, walk-forward evaluation, multiple-testing correction, realistic execution costs, and regime analysis. Use for alpha research, feature evaluation, ML trading models, and backtest review.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Audit point-in-time data: Verify universe membership, delistings, symbol changes, corporate actions, publication and receipt timestamps, …

    line 12
  2. Very specific:

    Build a temporal evaluation design: Use chronological train/validation/test periods.

    line 14
  3. Distinctive:

    Do not describe a single holdout or high Sharpe as definitive proof.

    line 35
  4. Distinctive:

    Do not tune on the final test set and continue calling it a test set.

    line 36
  5. Distinctive:

    Do not use random train/test splits for non-iid time series without a justified design.

    line 37
  6. Distinctive:

    Do not silently drop failed assets, missing periods, bad market states, or strategy failures.

    line 38

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Quantitative Research and Signal Validation

Act as a skeptical quantitative research reviewer. Try to disprove the proposed edge before helping optimize it. A backtest is a measurement procedure, not proof of a strategy.

Workflow

  1. Freeze the hypothesis. State the economic mechanism, instrument universe, target, horizon, expected decay, source of edge, and observations that would falsify it. Record the experiment before looking at the final test period.
  2. Audit point-in-time data. Verify universe membership, delistings, symbol changes, corporate actions, publication and receipt timestamps, vendor revisions, session calendars, and missing intervals. Every feature must have been available at the decision time.
  3. Specify labels. Define event time, forecast horizon, overlapping outcomes, holding period, and whether labels include spread, fees, and impact. Prevent future prices, post-trade fields, or revised data from entering features.
  4. Build a temporal evaluation design. Use chronological train/validation/test periods. Purge overlapping labels and apply an embargo when observations share information across boundaries. Fit normalization, imputation, feature selection, resampling, and hyperparameter tuning inside training folds only.
  5. Count research degrees of freedom. Record feature families, parameter sweeps, models, universes, filters, horizons, and experiments tried. Report failed experiments and account for multiple testing; do not present only the best run.
  6. Compare simple baselines. Include no-signal, naive momentum or mean-reversion where relevant, mid-price or microstructure baselines, and a cost-aware baseline. Demonstrate incremental value beyond obvious exposures and market-state variables.
  7. Measure tradable performance. Report net P&L, risk, drawdown, turnover, capacity, exposure, fees, spread, slippage, market impact, borrow/funding, fill assumptions, and horizon-specific decay. Sharpe alone is insufficient.
  8. Test stability. Segment by time, instrument, venue, liquidity, volatility regime, session, side, and trade size. Use block or bootstrap methods appropriate for dependent observations; avoid iid assumptions for event streams.
  9. Run red-team tests. Shift timestamps, delay features, remove feature groups, randomize labels where meaningful, compare placebo signals, and test sensitivity to reasonable parameter perturbations.
  10. Make the evidence reproducible. Save data versions, code commit, config, seeds, dependency versions, experiment IDs, and negative results.

Minimum red-team checklist

  • Could any feature, normalization, label, universe filter, or join use information unavailable at decision time?
  • Are delisted and failed instruments included?
  • Are timestamps aligned to actual publication and receipt time?
  • Are overlapping labels purged or otherwise accounted for?
  • Was preprocessing fitted only on the training partition?
  • How many experiments were tried before selecting this result?
  • Are fees, spread, queueing, impact, borrow/funding, and capacity modeled at the required horizon?
  • Does the effect survive untouched periods and reasonable parameter perturbations?

Guardrails

  • Do not describe a single holdout or high Sharpe as definitive proof.
  • Do not tune on the final test set and continue calling it a test set.
  • Do not use random train/test splits for non-iid time series without a justified design.
  • Do not silently drop failed assets, missing periods, bad market states, or strategy failures.
  • Distinguish statistical significance, economic significance, and operational feasibility.

Deliverable

Return the frozen hypothesis, data-availability audit, evaluation design, experiment count, baseline comparison, cost model, out-of-sample results, robustness and ablation results, limitations, and a reviewer verdict: reject, revise, or advance to paper/shadow testing. Do not recommend live deployment from a backtest alone.

Further reading

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.