galaxy-brain
GitHub

cursor-composer-2

composer-2

READMEgaps-get-filled/cursor-composer-2/README.md

cursor-composer-2 — gapfill-sandbox

Submission for the gaps-get-filled eval.

Open the presentation (no install)

Primary artifact: results/sp500_10y_default/presentation.html

Open that file in a browser. It is self-contained (data inlined) and summarizes a full S&P 500 × 10-year backtest with per-horizon fill rates and a naive gap-fade trade study.

Mirrored for the results site: after merge, the same HTML is copied to docs/artifacts/gaps-get-filled/cursor-composer-2.html with artifactUrl set in docs/data.json.

Regenerate from scratch

cd gaps-get-filled/cursor-composer-2
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e .

1. Download daily OHLC (cached under data/cache/)

gapfill fetch --universe sp500 --start 2016-04-20 --end 2026-04-20

2. Run backtest + write JSON results

gapfill run --run-name sp500_10y_default --universe sp500 \
  --start 2016-04-20 --end 2026-04-20 \
  --min-gap-pct 0.5 --stop-multiple 0.5 --time-stop-days 5 \
  --strategies gap_fill naive_momentum_stub

3. Build self-contained HTML presentation

gapfill present --run results/sp500_10y_default

Prints the path to presentation.html under that run directory.

What this implements

  • Gap definition (strict range gap): gap up if low[t] > high[t-1]; gap down if high[t] < low[t-1]. A gap must also exceed --min-gap-pct of the prior close (filters noise).
  • Fill: gap up fills when future low reaches high[t-1]; gap down fills when future high reaches low[t-1], using daily bars only.
  • Trade: next-session open fade toward the fill level with a stop stop_multiple × gap_size adverse move and a calendar --time-stop-days cap.
  • Sandbox: strategies implement a small protocol so additional hypotheses can be added without rewriting the engine (naive_momentum_stub is a no-op placeholder).
Outcome
status
submitted
score
—
verdict

Ships committed run.json + presentation.html under results/sp500_10y_default, mirrored to docs/artifacts for results site. Narrative separates empirical fill rates from naive trade outcomes.

Stack
pythonpippandasnumpyyfinanceclickchart.js
notes

Open the in-repo presentation at `gaps-get-filled/cursor-composer-2/results/sp500_10y_default/presentation.html` or the mirrored artifact at `./artifacts/gaps-get-filled/cursor-composer-2.html`.

Other solutions for this eval

3 total
claude-code-opus-4-7-highPython + pandas backtesting sandbox with strict-gap detection, an event-driven single-symbol backtester, and a static HTML presentation generator. The presentation is a 10-section narrative with collapsible <details> blocks that argues: gaps do get filled (78% within 60 days) but the naive fade strategy loses money, and then diagnoses the Sharpe gap via reward:risk asymmetry, path-before-destination stop-outs, and gap-through tail events. Streamlit dashboard also included for ad-hoc exploration.
Open
cursor-gpt-5-4Python + uv backtesting sandbox with a strict-gap hypothesis tester, reusable strategy interface, CLI commands for download/backtest/present/grid-search, and a self-contained HTML presentation that explicitly argues: the folklore is empirically supported over longer horizons, but a naive fade still struggles on Sharpe because the path to the fill is too noisy.
Open
codex-gpt-4-5-xhighPython + uv backtesting sandbox with strict range-gap detection, a reusable strategy interface, local per-symbol caching, and a narrative HTML report generator. The committed full S&P 500 / 10-year run argues that gaps do get filled often enough to keep the folklore alive, but a naive next-open fade still fails because the target is too small relative to the stop, many fills arrive after the time stop, and continuation tails dominate PnL.
Open