Improve and Rerun a Factor

Diagnose one tested factor and run a user-approved controlled experiment.

A failed or fragile factor is useful evidence. The safest improvement workflow diagnoses one exact run, changes one mechanism at a time, and preserves the original result.

1. Analyze The Exact Run

Use Factor Analysis for read-only diagnosis:

/quandora:factor-analysis analyze my latest factor result

If more than one run could match, select the exact factor and run instead of letting the agent guess. Factor Analysis reads server-persisted IS evidence and should separate:

• observed metrics and Health fields;
• inferences about the mechanism;
• plausible alternative explanations;
• controlled experiments that could distinguish them.

It does not submit, resume, archive, or modify anything.

2. Choose One Controlled Change

Prefer a change with a clear reason and measurable expected effect. Examples:

• reduce missing output while preserving the economic idea;
• smooth a noisy signal and check the autocorrelation/turnover tradeoff;
• change normalization to improve cross-sectional ranking;
• remove one weak input instead of adding several new inputs;
• test whether a filter causes an applicability/Health mismatch.

For each proposal, record the motivating evidence, expected change, metric that would confirm it, and main failure mode.

3. Confirm Before Creating A New Run

Analysis remains read-only. When you choose an experiment, explicitly ask Factor Mining to create and submit it:

Use Quandora Factor Mining to run the first proposed experiment. Keep the
original result unchanged and show me the exact change before submission.

The agent should show the new thesis and source-level change, validate the full new plugin.py, and obtain confirmation before submission.

4. Compare Like With Like

Compare runs only when their task, horizon, evidence scope, active-universe basis, and Health thresholds are compatible. Review:

• Success failure reasons;
• Health and missingness;
• Rank IC, autocorrelation, and Sharpe;
• path stability, drawdown, and turnover;
• whether the intended mechanism improved;
• whether a different risk appeared.

History And Immutability

Factor Mining can browse caller-owned factor families, versions, and runs. Historical source editing is not exposed in the current public agent workflow. A terminal result stays unchanged; an approved experiment creates new evidence instead of rewriting the old run.

Stop when repeated changes become curve fitting or when the evidence no longer supports the original mechanism. Continue to Strategy Construction only after you deliberately choose a factor for a strategy experiment.