Baz Cosmopoulos

What this is

A Deep RL agent that hedges a 10,000-share SPY position with liquid long-dated ATM calls & puts. The policy is Recurrent PPO (LSTM) trained on 100k rBergomi rough-volatility paths (GPU), then validated in Backtrader on historical SPY underlying + options. Goal: minimize PnL volatility and drawdown with materially lower cost than classical delta-hedging.

How it works

Simulator: rBergomi paths + ATM option prices (CuPy/CUDA).

Policy: RecurrentPPO (LSTM).

State (13 dims): SPY, ATM call/put mids, current call/put holdings, portfolio Δ & Γ, time-to-go, vol & lagged changes.

Actions (continuous, 2): call_qty ∈ [−100,+100], put_qty ∈ [−100,+100] per step (clipped/rounded to contracts).

Reward: minimize |ΔPnL| (or MSE/CVaR) + λ × transaction_costs; position limits enforced.

Training: HPO with Optuna, then long final train; evaluation over multiple seeds with VecNormalize.

Backtesting setup

• Engine: Backtrader (daily).
• Data: data/spy_underlying.csv, data/spy_options.csv (real options, ATM selection with fallbacks).
• Costs: $0.65/contract commission + 10 bps slippage on notional; contract multiplier 100; rolling near-expiry; daily rebalance.
• Constraints: ±200 contracts/type outstanding; ±100 contracts/trade; 30-calendar-day tenor target.

Results (vs classical delta-hedge)

• Volatility: −23.8% (0.35 vs 0.46).
• Max drawdown: −36.3% (−0.85 vs −1.33).
• Trading costs: −96.0% ($83,168 vs $2,087,810).
• Hedging efficiency: 12.02× (vs 0.48).
• Win rate across seeds: 95%.

(See repo tables for full metrics: downside deviation, ulcer/pain indices, equity stability.)

Why this matters (for trading)

• Frames hedging as a cost-aware control problem (risk + frictions), not “hedge everything, every step.”
• Produces an explicit risk/cost Pareto surface you can choose from based on liquidity and risk budget.
• Same idea generalizes to inventory-aware execution / market making: state → action → costs, with constraints.

Failure modes & gotchas

• Sim-to-real: rBergomi misspecification, discrete strikes, and real microstructure can break “optimal” actions.
• λ sensitivity: mis-setting λ can under-hedge (cheap but risky) or over-hedge (stable but too expensive).
• Liquidity shocks: wide markets and gap risk can push slippage beyond the backtest cost model.
• Ops details: contract rounding, roll/assignment, missing chains, and stale quotes need robust fallbacks.

Reproduce quickly

pip install -r requirements.txt && pip install -e .
# Hyperparameter search (default 10 trials)
hedgerl train --hpo
# Final training with your chosen loss/weights
hedgerl train --loss_type abs --w 0.05 --lam 0.001
# Backtest (single or looped seeds)
hedgerl backtest
hedgerl backtest --loop 5
# Explore Pareto frontier (risk ↔ cost)
hedgerl generate-pareto

Repo map (just the bits recruiters care about)

src/sim/rbergomi_sim.py (GPU rough-vol paths & ATM pricing) • src/env/hedging_env.py (state/action/reward) • src/agents/train_ppo.py (RecurrentPPO + Optuna) • src/backtester/* (Backtrader, model wrapper, delta baseline) • model_files/* (exported weights + normalization)

Notes & limitations

Sim-to-real gaps (liquidity, microstructure) remain; ATM selection uses robust mids with BS fallback; commission & slippage are parameterized; single-asset scope (multi-asset on roadmap).