← Blog/blog/option-hedging-calendar-blocks

The 3.1% tail-risk win that one market path cannot certify

An options market maker inherits directional exposure when clients trade. Textbook delta hedging cancels the first-order price sensitivity with stock. Chen and colleagues instead let a reinforcement learning agent choose a hedge ratio after seeing the index level and the option book's delta.

The headline is attractive: train on real S&P 500 option history and the worst 5% of outcomes improve by 3.1% versus delta hedging. But those outcomes are overlapping windows cut from one realized market path. The uncertainty calculation—not the neural network—is the post's spine.

01

What the agent is actually choosing

Let the client book have delta δ. A delta-neutral desk holds−δ shares of the index. The agent outputs a scalar hedge ratioa and holds −aδ shares, so net delta is(1−a)δ. With a negative-delta book of −40 and the paper's approximate average ratio 0.95, the desk remains at -2.0net delta: a small persistent short tilt.

CVaR95, or 95% conditional value-at-risk, is the mean profit and loss inside the worst 5% tail. It answers “how bad is an average disaster?” rather than “what is the average day?” The TypeScript below is the tested code used for the calculations; Python and C++ are line-for-line translations.

def empirical_cvar(pnls, confidence):
    if not pnls or not 0 < confidence < 1:
        raise ValueError('invalid sample or confidence')
    tail_count = max(1, math.ceil((1 - confidence) * len(pnls)))
    tail = sorted(pnls)[:tail_count]
    return sum(tail) / len(tail)

def net_delta(book_delta, hedge_ratio):
    stock_holding = -hedge_ratio * book_delta
    return book_delta + stock_holding
02

The gain is regime-dependent insurance

The empirically trained distributional TD3 agent improves tail losses in falling markets and pays for that protection in rising markets. Its state-dependent short tilt is therefore closer to crash insurance than a free reduction in every kind of risk.

Paper-reported CVaR95 loss magnitudes. Lower is better. The empirical RL policy helps most in the high-volatility falling regime and is worse in both rising regimes.
RL minus delta hedging
Core-computed percentage change in tail loss from the paper's regime table, ordered Q1 through Q4. Positive favors RL; negative favors delta hedging.
Hedge ruleOverall CVaR95Q2 fallingQ4 risingMean PnL
Constant ratio 0.97+1.1+4.9−0.0+0.21
Short offset 5+1.1+8.7−2.7−0.81
Band around 0.95δ+1.3+7.5−0.5+0.41
Empirical TD3+2.5+17.9−3.8−0.65
Paper-reported changes versus delta hedging. Positive tail-risk changes favor the rule; positive mean PnL favors the rule. Simple rules recover part, but not all, of the learned policy's crash protection.
03

The bootstrap decides what can be claimed

A bootstrap resamples data to approximate repeated experiments. Treating 1,386 overlapping start dates as clusters preserves episode-level dependence but still acts as though the calendar supplied many independent stress histories. Resampling 63-day calendar blocks keeps longer stretches together and exposes how little independent tail evidence exists.

Bootstrap scheme95% interval (CVaR points)Includes zero?
Window-start clustered[+1.1, +3.8]No
63-day calendar blocks[−2.1, +8.1]Yes
Paper-reported intervals for the same 2.5-point CVaR95 improvement. The dependence assumption changes the inferential verdict.
Core-computed confidence-interval widths from the paper's reported endpoints. Calendar blocking makes the uncertainty range almost four times wider.
04

Where the replay experiment can mislead

The client-flow process is synthetic and directionally symmetric; the observation omits the desk's carried stock position; every episode is only 40 days; stock costs are fixed at 3 basis points; financing is zero; and live options at the horizon incur no liquidation cost. These choices hold comparisons together, but they are not a production market-making simulator.

More subtly, the RL policy is memoryless even though the environment is only partially observed. A robust next test would use multiple exchanges or indices, rolling calendar holdouts, and a policy that sees hedge inventory and funding costs. Report both calendar-block uncertainty and regime-specific utility before calling a tail-risk edge durable.

05

What to probe next

Sweep the hedge ratio and transaction cost together, then compare the learned policy with a simple state-dependent short-offset rule on truly separate market histories. The site's neural networkpage helps unpack the actor and quantile critics; thelinear model page is the right baseline for a transparent regime-dependent hedge.

References

  1. Chen, J. et al. (2026). Risk-Sensitive Reinforcement Learning for an Options Market Maker: An Empirical Study with Real S&P 500 Options. arXiv:2611.10956
  2. Rockafellar, R. T., and Uryasev, S. (2000). Optimization of Conditional Value-at-Risk. Journal of Risk 2(3)
  3. Buehler, H., Gonon, L., Teichmann, J., and Wood, B. (2019). Deep Hedging. Quantitative Finance 19(8)