Institutional-grade execution-analytics figures

[joshfinney]

What

Replaces the previous ad-hoc, dark-themed / radar-style plots with a clean, buy-side-ready figure suite generated directly from the eval data. The old charts were never committed; this adds a proper, reproducible plotting module.

Why the old plots read as "vibe-coded"

• Dark neon deck — institutional/TCA reports are light and print-friendly.
• Radar/spider charts — a known anti-pattern for quantitative comparison (misleading area, arbitrary axes, unreadable overlapping fills) that also hid the actual bps numbers.
• Saturated rainbow palette, colored annotation text, chartjunk, no hero/recessive discipline.

What's here now

Five figures (experiments/plots.py), all in basis points, light theme, one accent for the hero policy (PPO), muted-gray baselines, with markers + line styles as secondary encoding (survives greyscale / CVD):

1. Execution efficiency frontier — cost vs. risk (the signature Almgren–Chriss view)
2. Implementation-shortfall distribution — box/whisker + 95% VaR marker
3. Mean cost by regime — grouped bars with ±1σ error bars
4. 95% VaR tail-risk league table — ranked horizontal bars
5. Execution schedule — remaining inventory over the horizon

Palette validated with the dataviz colour checker (CVD separation ΔE 38.5, well clear of the floor).

Story the charts tell

TWAP posts the lowest mean cost, but PPO delivers the best risk-adjusted execution — roughly half the cost dispersion and the lowest tail risk (VaR) in every regime.

Structure

• viz/style.py — shared theme (palette, rcParams, bps formatter)
• experiments/collect_raw.py — per-episode costs + trajectories
• experiments/plots.py — figure generation (PNG for README + SVG vector)
• eval.py --figures flag; README Results section with embedded charts

Reproduce: uv run python experiments/plots.py. Tests still green (8 passed); ruff clean.

🤖 Generated with Claude Code