gamma-scalper

Short gamma scalping on Deribit BTC/ETH options, delta-hedged via perpetuals.

Edge: IV - RV premium (+6.5% mean, 72% win rate) + funding income (+5.3% ann). Calibrated on 730 days of BTC data (Jun 2024 - Jun 2026).

---

Architecture

config/
  market.toml       venue constants, fees, tick sizes
  strategy.toml     AS model, vol premium signal, sizing, rolling, OFI
  risk.toml         position limits, drawdown, kill switch triggers
  execution.toml    order types, cancel/replace, latency budgets
  loader.py         Pydantic validation + hot-reload

core/
  state_engine      L2 book + OFI, Yang-Zhang RV estimator, delta tracker,
                    funding regime, SVI + SABR vol surface
  strategy          AS pricer (SABR vanna-adjusted), MLE k calibration,
                    OFI entry filter, multi-leg position registry,
                    straddle/strangle support, roll detector,
                    regime-conditional Sharpe filter, calendar spread hedge
  execution         order lifecycle, simultaneous roll, iceberg support,
                    smart partial fills, queue position estimation
  risk_engine       real fill PnL, per-leg attribution, live margin from
                    exchange, Telegram/Slack alerting, kill switch
  market_data       Deribit WS feed, SVI/SABR refit on greeks, resync

infra/
  deribit_gateway   JSON-RPC auth, push fill notifications (no polling),
                    private WS reconnect with re-auth, Black-76 pricer,
                    account summary for margin monitor
  logging_setup     structured JSON logs (Datadog/Loki/Grafana), StatsD
                    gauges, rotating file handler

tests/
  test_integration  14 end-to-end tests, no real WS needed
main.py             asyncio orchestration, dry-run mode, graceful shutdown

---

Quickstart

pip install -r requirements.txt

# signal flow only, no orders
python main.py --dry-run

# testnet (use_testnet=true in config/market.toml)
export DERIBIT_CLIENT_ID=your_id
export DERIBIT_CLIENT_SECRET=your_secret
python main.py

# live
# set use_testnet=false in config/market.toml, then:
python main.py

---

Config

Nothing hardcoded. Every threshold, model parameter, and fee lives in config/.

File	Contents
market.toml	Venue URLs, fee tiers, tick sizes, book validation
strategy.toml	AS model, vol premium, OFI filter, sizing, rolling, hedge vehicle
risk.toml	Position limits, drawdown windows, kill switch triggers
execution.toml	Order types, cancel/replace triggers, iceberg, latency budgets

Key parameters calibrated from 730-day BTC dataset:

# strategy.toml
[vol_premium_signal]
entry_threshold          = 0.05    # IV - RV > 5% to enter
emergency_exit_threshold = -0.15   # flatten below -15%

[funding_regime]
size_multiplier_bull    = 1.0      # >5% ann funding: full size
size_multiplier_neutral = 0.7
size_multiplier_bear    = 0.3      # negative funding: stay small

[ofi]
entry_threshold = 0.60             # skip entry if |OFI| > 0.6

[delta_hedge]
delta_threshold = 0.05             # hedge when accumulated delta exceeds this

[realized_vol]
estimator = "yang_zhang"           # YZ: ~5-8x more efficient than C2C

# risk.toml
[kill_switch]
rv_spike_halt_threshold   = 3.0    # RV(1h)/RV(24h) > 3x: flatten
funding_negative_halt_ann = -0.20  # funding < -20% ann: flatten

Config hot-reloads every 5 minutes without restart.

---

Strategy

Entry conditions (all must hold):

• IV - RV > entry_threshold (default 5%)
• |OFI| < 0.60 (no strong directional flow on the perp)
• Rolling 30-day Sharpe of vol premium > sharpe_filter_threshold
• Funding regime multiplier > 0 (not in confirmed bear regime)

Pricing: Avellaneda-Stoikov in vol space. Reservation price adjusted for inventory skew and SABR vanna (how much IV moves with spot). AS arrival rate k calibrated via Poisson MLE from live fill data.

Structures:

• straddle: sell ATM call + put, same expiry
• strangle: sell OTM call + put at configurable delta targets (e.g. 25-delta)

Hedge vehicle:

• Default: BTC-PERPETUAL (funding income when positive)
• Fallback: quarterly futures when funding is persistently negative (calendar spread)

Rolling: Three independent triggers: DTE < threshold, moneyness drift > 5%, vol surface shift > 10%. Simultaneous roll — close and open fire concurrently via asyncio.gather, no gap between legs.

Multi-asset: ETH config is included. Add "ETH" to active_assets in strategy.toml and the coordinator spins up a second feed + strategy loop automatically.

---

Kill switch triggers

Any one fires an immediate flatten + halt + alert:

• RV spike > 3x (1h vs 24h)
• Funding < -20% annualized
• Intraday drawdown > $2,000
• 24h drawdown > $3,000
• Loss velocity > $500/h or $100/min
• Margin utilization > 80%
• Perp/index divergence > 5%
• Book stale > 5s
• 3 consecutive API errors
• Private WS silent > 10s

---

Observability

Logs: structured JSON to stdout + rotating file. Compatible with Datadog, Grafana Loki, and any collector that reads JSON lines.

Metrics: StatsD gauges emitted from every risk snapshot. Set STATSD_HOST to enable.

Alerts: Telegram and/or Slack on halt. Set TELEGRAM_BOT_TOKEN + TELEGRAM_CHAT_ID and/or SLACK_WEBHOOK_URL.

---

Running tests

python tests/test_integration.py
# 14 passed, 0 failed

---

Environment variables

# required
DERIBIT_CLIENT_ID
DERIBIT_CLIENT_SECRET

# optional - alerting
TELEGRAM_BOT_TOKEN
TELEGRAM_CHAT_ID
SLACK_WEBHOOK_URL

# optional - metrics
STATSD_HOST          # e.g. localhost
STATSD_PORT          # default 8125