Automated analysis

This analysis was drafted by our research engine and has not been checked by a human editor. It may contain errors. It separates the paper’s own results from our tests, and any figures called ours come from our own backtest.

Our automated analysis

Supervised exits triple 2024-2026 SPX short-straddle returns

Actor-Critic adds 2.53 return points over the classifier (p=0.184) across seven seeds, and 0.45 with one-contract sizing.

2026-10-12 · 5 min read · Options exit timing / reinforcement learning · SPX index options (no feasible intraday implementation)

Reviewing: MLP-Initialized Reinforcement Learning for Exit Timing in SPX Short-Straddle Portfolios · Jianning Wang, Robert Ślepaczuk and Paweł Sakowski · Read it on arxiv

A short SPX straddle book did far better with supervised exits than with calendar or stop rules. The extra return from reinforcement learning is much harder to credit.

What did Wang, Ślepaczuk and Sakowski train?

The trade draws on the variance risk premium: index implied vol tends to exceed realized vol, paying straddle sellers for bearing large moves. Wang, Ślepaczuk and Sakowski take that carry as given. Their question is when to buy back a straddle already on the book.

The strategy opens one SPX straddle each day, roughly 20 trading days from expiry, selling at the bid. A 20-slot rolling portfolio starts with $3,000,000. After a forced 26-bar hold, about two trading days, the policy can choose hold or exit every 30 minutes, up to 13 times daily. It buys back at the ask and pays $1.50 per contract per leg in commissions. Its 17 contemporaneous features include relative bid-ask spread, a VRP proxy, remaining time value as a share of the mid, unrealized PnL and the SPX move since entry.

Training begins with hindsight labels. For each 2016-2023 episode, the authors identify the ex-post best exit and label every subsequent bar "exit"; a two-hidden-layer MLP learns those labels. They then transfer its weights to Actor-Critic, advantage-weighted regression (AWR) and AWAC actors for fine-tuning. The shaped reward pays mark-to-market change and a VRP holding bonus, while charging spread on exit and penalties for trailing CVaR and late holding. Results come from a locked 2024-2026 window of about 2.42 years.

Mechanical exits lag the classifier

Out of sample, the supervised MLP returns 35.01% with a Sharpe of 1.651. Hold-to-expiry returns 11.56% with a Sharpe of 0.441. Both use the paper's compounding stress-budget sizing. Under one-contract sizing, returns are 35.04% against 15.67%, about 2.2 times rather than three. Maximum drawdown drops from 12.82% to 4.04%; the longest underwater stretch shrinks from 0.73 years to 0.29.

The mechanical rules fare worse than holding to expiry. A fixed 65-bar exit loses 2.12%. A 25% profit target paired with a 50% stop earns 1.91%, while the combined rule earns 1.15%.

The MLP wins by cutting losses, despite a 56.2% hit rate against hold-to-expiry's 56.9%. Its drawdown is a third of hold-to-expiry's. Early exits also free slots: the MLP enters 559 trades against 425, so capital reuse contributes to the gain. Episode PnL suggests it is a secondary source. In the one-contract run, the MLP takes 577 trades against 454, yet mean episode PnL is $1,803 against $818 for holding to expiry.

The benchmark choice matters here. The authors call the mechanical rules "representative", but give each a single parameter choice. Before reading 35% versus 2% as the general gap to mechanical exits, I would want a small grid of holding periods and thresholds.

How much does Actor-Critic add?

Across seven seeds, MLP-initialized Actor-Critic averages 37.54% ± 0.92% and a Sharpe of 1.705. The gains over the MLP are 2.53 return points and 0.054 of Sharpe. The authors concede that the episode-level difference is insignificant: mean paired gain is $129.94 (t=1.330, p=0.184), with Wilcoxon p at 0.681. They defend an improvement that remains "positive on average" in economic terms.

The seed choice weakens that defence. Headline runs use the seven odd seeds 17 through 29. In the reward ablation, the full shaped reward gets 35.56% ± 2.73% on seeds {1,3,5,7,11}, just 0.55 points above the MLP. The paper explains why the ablation arms share seeds {1,3,5,7,11}, without explaining the change from the headline set. Across five seeds, AWR earns 35.23% and AWAC 35.19%. Both sit within 0.22 points of the MLP's 35.01%, with seed standard deviations of 1.93 and 1.76 points.

The AC+MLP ordering survives one-contract sizing, as the authors say, by 0.45 points. Switch off dynamic sizing and AC+MLP returns 35.49% ± 1.93% against the MLP's 35.04%, with p=0.645. That 0.45-point lead lies inside a 1.93-point seed standard deviation. The authors describe the increment as "amplified by dynamic position sizing". Their sizing rule compounds contract count from slot NAV, letting a good early run buy a larger book later.

Reward shaping offers little further support. Removing the CVaR penalty raises return slightly, to 36.80%. A direct reward without the VRP bonus or CVaR term produces 35.11%. The coefficients were fixed in advance and barely shift the result.

Actor-Critic trained from random weights returns 0.18% ± 4.35%, with a Sharpe of -0.102. The failure is real (paired p=3.01e-13). Yet the configuration table assigns the MLP-initialized run a BC coefficient (presumably behaviour cloning) of 0.05 and the scratch run 0.00. The paper says its RL variants differ only in initialization and policy-update objective. Here regularization changes too, weakening the conclusion that supervised initialization makes reward-based fine-tuning viable.

Hindsight earns 373%

Through the same simulator, the hindsight oracle returns 373.02% and earns $19,261 per episode on average. The MLP earns $1,802. Its labels therefore target more than ten times the classifier's result (10.7x). The MLP captures about a tenth of oracle per-episode PnL ($1,802 of $19,261), which the authors call "only a limited fraction".

The floor is the bigger worry.

The authors warn that the roughly 4% drawdowns "should therefore not be interpreted as upper bounds on naked short-straddle tail risk". The 2024-2026 test contains nothing like 2008 or March 2020. Design adds another constraint: during the 26-bar minimum hold, no policy can close a position in its first two days. Only maintenance-margin liquidation can intervene.

Why we could not replicate the 30-minute exits

Our option data is end-of-day only. This policy needs matched call and put bid and ask quotes at every 30-minute bar to calculate relative spread and time value, then fill exits at the ask. Faithful replication requires intraday SPX option quotes for both legs, stamped alongside the index level, across 0.95 to 1.05 moneyness and 0 to 21 days to expiry.

The supervised exit model is the part worth acting on; the RL layer is optional. I would change my view of the MLP after a retrain ending before 2020 and a test through that March. March 2020 appears in the training labels, never in a test. If drawdown stays near 4% there, its loss truncation is real.