gex-strategy

Long-only GEX strategy for US equities. Trades dealer gamma/vanna positioning from options data, enters on IV spike → compression reversals, exits via bracket SL/TP + wall-trailing stop. Rust backtester + live engine with IBKR bracket orders and React dashboard.

Not financial advice. Educational/research only. Trading = risk. Past performance ≠ future results.

Want to help? I'm looking for people to stress-test this: find lookahead bias, break the half-period stability, challenge execution assumptions, or improve entry gate logic. Run --overfit-report or --missed-entries and tell me what's wrong. See the full ask below.

---

Quick start

Need Rust and Git LFS.

git clone https://github.com/simonrey1/gex-strategy.git
cd gex-strategy
git lfs pull  # ~315 MB parquet data (8 years, 15 tickers)

cargo run --release --bin backtest                           # full portfolio
cargo run --release --bin backtest -- --ticker AAPL -p 8080  # single ticker + dashboard

What you get (~30 seconds after first compile):

[backtest] === RESULTS: AAPL,GOOG,MSFT,JPM,GS,WMT,HD,DIS,KO,CAT,MS,NRG,SO,MCD,COST ===
  Period:       2018-01-15 -> 2026-04-01
  Bars:         50766 (15-min)
  Trades:       493
  Win/Loss:     217W / 276L  (44.0%)
  Gross PnL:    $132607.43
  Commission:   -$1031.21 (~$2.09/trade)
  Slippage:     -$9430.20
  Net PnL:      $122146.02
  ┌───────────────────────────────────┐
  │  Strategy:   +1221.46% on $10000  │
  │  Buy & Hold: +235.59%             │
  │  Alpha:      +985.87%             │
  └───────────────────────────────────┘
  Profit Factor: 2.99
  Max Drawdown: $6129.31 (9.03%)
  Sharpe:       2.82
  Sortino:      4.94
  Calmar:       3.93
  CAGR:         35.47%
  Ulcer Index:  0.0192
  Expectancy:   $247.76/trade  |  Payoff: 3.44
  Max DD Dur:   67 trading days  |  Avg Hold: 13395 min  |  Avg Capital Util: 49.2%
  Avg Win:      +6.24%  |  Avg Loss: -1.81%  |  Avg Trade: +1.73%

For the dashboard (needs Node.js):

cargo test --release export_bindings   # generates TypeScript bindings
cd dashboard && npm install && npm run build && cd ..
cargo run --release --bin backtest -- --ticker AAPL -p 8080

Opens http://localhost:8080: candlestick chart with GEX walls, trade markers, P&L curve, IV analysis.

No API keys, no subscriptions. Everything runs offline in ~30s.

cargo run --release --bin backtest -- --start 2022-01-01 --end 2024-01-01   # date range
cargo run --release --bin backtest -- --sweep "vf_min_atr_pct=0.20,0.25,0.30,0.35"  # param sweep
cargo run --release --bin backtest -- --overfit-report                       # stability analysis
cargo run --release --bin backtest -- --help                                 # all flags

---

Thesis

Exploits dealer hedging flows from options GEX walls. Needs deep options liquidity: high OI creates meaningful walls that act as support/resistance. Works on the top ~20 US equities by options OI (AAPL, GOOG, MSFT, JPM, GS…). Same names dominate year after year.

Signals

VannaFlip (VF) — the main signal

When IV spikes — options fear surges — dealers who sold puts hedge aggressively. When the panic fades and IV compresses, those same dealers unwind: they buy back stock, creating a mechanical tailwind. The strategy detects the storm, then waits for the calm before entering.

1. SPIKE DETECTION — "Is there a storm?"
   ├─ IV spikes to ≥3.5× its baseline (real panic, not noise)
   ├─ Price near the put wall (near support — where hedging pressure peaks)
   └─ Minimum volatility (market active enough to trade)
   → 50-bar window (12.5h) opens.

2. ENTRY GATES — "The calm after the storm"
   Checked every 15-min bar within the window. All must pass:
   ├─ IV Compression : IV dropped to ≤50% of spike peak (fear fading)
   ├─ Max ATR% < 0.50 : price choppiness settling (not just fear headline)
   ├─ Wall structure : GEX walls present and healthy
   ├─ Trend state : TSI, momentum, dead-zone checks
   └─ 5 more structural checks (GEX norm, wall spread, PW/CW strength)
   → First bar where everything passes: buy.

Uses today's peak IV (iv_daily_high), not the current reading. A 9:35am flash spike keeps the window open all day — the panic already happened.

IvCompress + AtrPct carry most of the Sharpe — they are the "calm after the storm" filters.

IV scan oracle (--missed-entries): after spike detection, calls IvCompressionInputs::is_eligible() (has spike + within 50 bars + IV exists), then enters at every eligible bar — skipping all VF entry gates. Classifies results as best/worst. The b/w ratio is affected by Stage 1 (spike detection) and is_eligible(), but NOT by VF entry gates. b/w ≈ 0.28 is invariant to spike-time conditions; improving it requires smarter filtering in is_eligible() (compression-time environment).

WallBounce (WB)

Calm-path entry near put wall via zone-dwell score. Currently JPM only.

Exits

Bracket SL/TP + wall-trailing stop + Hurst exhaustion trailing.

Portfolio

max_open_positions=3 across 15 tickers. rank_and_dedup sorts candidates by TSI, fills available slots. ~78% of signals get skipped because slots are full: this involuntary scarcity does most of the filtering. TSI ranking itself isn't predictive (inverting sort order gives similar Sharpe).

Regime risk

The edge depends on IV spikes. Low-vol sideways markets → fewer spikes → Sharpe roughly halves. No spike = no dealer flow to ride.

Robustness analysis

Overfit report has half-period stability, 4-period checks, per-ticker sweeps. Note: params were tuned on the full period, so H1/H2 is a stability check, not a true holdout.

Regenerate: cargo run --release --bin backtest -- --overfit-report

CLI flags

Flag	Default
--ticker <SYM[,SYM]>	all	Comma-separated
--start/--end <DATE>	2018-01-01 / 2026-03-01
-v <0-3>	1	0=sweep line, 1=summary, 2=trades, 3=rejections
-p <PORT>	off	Dashboard
--overfit-report		Half-period + 4-period stability report
--missed-entries		Best entries blocked by VF gates (+ dashboard mini charts)
--sweep "<p=v1,v2>"		Single-process param sweep

Dashboard

Served by backtest (-p 8080) and live (:3000).

• Candlestick + GEX wall overlays + trade markers
• IV panel: ATM put IV, baseline, spike events, compression points
• IV Scan: simulates hypothetical entry at every compression bar post-spike, classifies as Best/Worst
• Trade analysis: high-runup losers (ran up but reversed to SL) + worst losses

Live trading

Needs Podman, ThetaData Pro, IBKR account.

kube/play.sh dev                    # IB Gateway + Theta Terminal
cargo run --release --bin live      # strategy + dashboard on :3000

Production: kube/play.sh prod. See live trading docs for VPS setup.

---

Tests

cargo test && npm --prefix dashboard test

Non-regression snapshots:

INSTA_UPDATE=always INSTA_FORCE_PASS=1 cargo test --release --test portfolio_nonreg && cargo insta accept

Want to help?

Params were tuned on the full 2018–2026 period. H1/H2 is a stability check, not a true holdout. I want people to find what's wrong.

• Find lookahead bias in entry signals, GEX wall computation, IV baseline, or exit logic
• Break half-period stability: find a param that pumps H1 Sharpe while H2 drops. --overfit-report does this but might miss subtler cases
• Challenge execution: are commissions ($0.005/share), slippage ($0.10), fills realistic?
• Improve entry gates: --missed-entries -p 8080 shows the best entries that gates blocked, with mini charts. Better gate logic > param tweaking

Some params are sharp peaks (fragile?): exit_width_atr, spike_min_atr_pct, vf_max_atr_pct all drop Sharpe significantly with ±0.05 changes. Real structure or noise?

License

CC BY-NC-SA 4.0, non-commercial use only.

Docs

Doc
Strategy	Signals, SL/TP, indicators, risk
Architecture	Data flow, types, GEX walls, cache
Backtest	Execution, commissions, chart markers
Live Trading	Poll loop, IBKR orders, VPS deployment
Robustness	Half-period sweeps, per-ticker analysis