report-research

Overview

This skill turns an experiment run into a durable write-up: a full experiment.md plus the standard show_experiment plot(s) linked from the report.

Workflow (do all steps)

1) Locate the experiment folder

Use the folder that contains:

• configs/ (the configs you ran)
• results/ (JSON metrics output)
• predictions/ (OOF parquet output)
• experiment.md (the report you will write/update)

2) Inventory what was actually run

• List configs that exist.
• Determine which ones were executed by checking for matching results/*.json and predictions/*.parquet.
• Identify the “best” model(s) using bmc_mean and bmc_last_200_eras.mean (primary), with corr_mean as a sanity check.
• If experiments were run in rounds, summarize each round’s intent (what changed) and whether it improved the current best.

3) Extract metrics for the report

For each run you report, include at least:

• corr_mean
• bmc_mean
• bmc_last_200_eras.mean
• avg_corr_with_benchmark (from the BMC summary)

Prefer a single markdown table with one row per model.

4) Write a full report in experiment.md

Update/create experiment.md with these sections (keep it crisp but complete):

• Title + Date
• Abstract (what was tested + headline result)
• Hypothesis / Motivation (why this should help BMC)
• Method (data split, CV, feature set, model type/hparams, any transforms)
• Experiments run (one subsection per config that actually ran; include output artifacts)
• Results (the metrics table; mention best run + trade-offs)
• Standard plot (embed the PNG and include the generating command)
• Decisions made (what you chose and why; e.g., per-era vs global, feature set choice, sweep choices)
• Stopping rationale (why you stopped iterating; e.g., plateau after N rounds, confirmatory scale step, diminishing returns)
• Findings (what worked / didn’t; interpret the plot)
• Next experiments (2–5 concrete follow-ups)
• Repro commands (train + plot commands from repo root)

5) Generate the standard plot(s) and link them

Default standard plot (baseline = benchmark predictions):

PYTHONPATH=numerai python3 -m agents.code.analysis.show_experiment benchmark <best_model_results_name> \
  --base-benchmark-model v53_lgbm_ender20 \
  --benchmark-data-path numerai/v5.3/full_benchmark_models.parquet \
  --start-era 575 --dark \
  --output-dir numerai/agents/experiments/<experiment_name> \
  --baselines-dir numerai/agents/baselines

Then embed it in experiment.md with a relative link:

![benchmark vs best model](plots/<generated_plot_name>.png)

If you have multiple candidate models, either:

• generate one plot with multiple experiment models, or
• generate one plot per candidate (and link all of them).

6) Final checks

• Plot files exist under plots/.
• experiment.md links resolve (use relative paths).
• Metrics table matches results/*.json.
• Report clearly states what was run vs what is only planned/configured.