gpt-5-4-backtesting

Submission for the gaps-get-filled galaxy-brain eval.

👉 Open the presentation right now

In-repo HTML: results/sp500-10y/presentation/report.html

Clickable hosted mirror: galaxybrain.dev/artifacts/gaps-get-filled/cursor-gpt-5-4.html

The presentation takes a clear position:

Strict price gaps do often get filled over time, but a naive gap-fade trade is still not a good risk-adjusted strategy.

It shows both sides of that sentence explicitly:

• a per-horizon fill-rate study at 1d / 5d / 20d / 60d
• actual trade results for a configurable fade strategy
• a simple Sharpe-style diagnostic explaining why "empirically true" does not automatically mean "tradeable"

Stack

• Python + uv
• pandas and yfinance
• static Plotly HTML presentations
• a reusable strategy sandbox so later ideas slot in cleanly

Regenerate from scratch

uv sync

# fetch/cache daily price data for the S&P 500
uv run gpt-5-4-backtesting download \
  --universe sp500 \
  --years-back 10

# run the baseline presentation build
uv run gpt-5-4-backtesting present \
  --universe sp500 \
  --years-back 10 \
  --min-gap-pct 0.5 \
  --max-hold-days 5 \
  --stop-gap-multiple 0.5 \
  --report-dir results/sp500-10y/presentation

# sweep a few nearby variants
uv run gpt-5-4-backtesting grid-search \
  --universe sp500 \
  --years-back 10 \
  --min-gap-pcts 0.5,1.0,1.5,2.0 \
  --max-hold-days-grid 3,5,10 \
  --stop-gap-multiples none,0.5,1.0 \
  --sort-by avg_return_pct \
  --min-trades 50 \
  --report-dir results/sp500-10y/grid-search

The present command prints the HTML path and produces a self-contained presentation you can open directly in a browser. The grid-search command produces a second HTML companion report showing how small parameter tweaks change the result.

Hypothesis definition

This submission uses a strict gap definition:

• gap up: today.low > yesterday.high
• gap down: today.high < yesterday.low

A gap is counted as filled once price later trades back through the prior bar's extreme:

• gap up fills when price trades back down to yesterday's high
• gap down fills when price trades back up to yesterday's low

That empirical statement is tested separately from the actual trade rule. The default strategy:

1. detects a strict gap
2. enters contrarian at the gap-day open
3. targets the prior bar boundary
4. exits on fill, stop, or time stop

That split is intentional: the whole point of the presentation is to show how a claim can be true in the data while still not producing a good Sharpe.

CLI overview

uv run gpt-5-4-backtesting download --universe sp500 --years-back 10
uv run gpt-5-4-backtesting backtest --universe sp500 --years-back 10
uv run gpt-5-4-backtesting present --universe sp500 --years-back 10
uv run gpt-5-4-backtesting grid-search --universe sp500 --years-back 10

Every command also accepts --symbols AAPL,MSFT,NVDA instead of a named universe and supports either --years-back or explicit --start / --end.

Repo layout

src/gpt_5_4_backtesting/
  cli.py          CLI entry points
  data.py         daily OHLCV loader + cache
  universes.py    named universe resolution (currently `sp500`)
  engine.py       backtest + fill-rate study + summary metrics
  grid_search.py  small parameter sweep runner
  reporting.py    static HTML presentation builders
  strategies/
    gap_fill.py   strict-gap detection + baseline fade strategy
tests/
  test_gap_fill.py
  test_grid_search.py
results/
  sp500-10y/
    presentation/
    grid-search/

Extending it

The code is structured so future strategies can reuse the data layer, metrics, CLI plumbing, and presentation pattern. The next obvious experiments are:

• split gap-up and gap-down trades
• filter by market regime
• exclude earnings gaps
• replace trade-level Sharpe with a portfolio-style daily equity curve Sharpe