Predictive Gamma Strategy

Investment objective. Extract intraday short-dated premium on the S&P 500 by trading 0DTE straddles only on days the model identifies a high-conviction edge, and standing aside otherwise. The edge claimed is selectivity — when not to trade.

Instruments / venue. 0DTE (same-day expiry) S&P 500 (XSP) options, traded as an at-the-money straddle (call + put at approximately the at-the-money strike), long or short depending on the signal. Listed, cash-settled, European-style; settled at the close.

ML model description (methodological). A gradient-boosted decision-tree classifier trained on a rolling recent window of trading days. Each day the model is re-fit on the most recent window and predicts the probability that selling that day's at-the-money straddle would be profitable; the training label is whether a same-day short straddle would have been profitable historically. Inputs are engineered from the volatility complex and recent price/volatility behaviour and include, but are not limited to, short-dated-versus-standard implied-volatility relationships (for example, a short-horizon VIX measure relative to the standard VIX), VVIX momentum, and realised-versus-implied move spreads. Because it is retrained continuously on rolling data, the model adapts to the recent regime rather than relying on a single static training.

Signal interpretation guide. The model produces one probability per day, mapped to three postures:

Short straddle — probability of a profitable short above an upper threshold (model anticipates a "volatility crush"/quiet session).
Long straddle — probability below a lower threshold (model anticipates a "volatility breakout"/large move).
Flat (no trade) — in the dead-band between the thresholds; low-conviction days are skipped by design.

A signal expresses a probabilistic lean, not a certainty — high-conviction days can still lose, and a 0DTE straddle carries severe intraday risk (see Risk Disclosure, §4.4).

Position sizing. Fixed and NAV-proportional. No other targeting.

Rebalancing triggers. Daily / intraday: a posture is taken (or skipped) each session; positions are 0DTE and expire/settle the same session; the next day's posture is re-evaluated from scratch. There are no multi-day rolls.

Why the benchmark differs. Strategies 1–3 are overlays on a core S&P 500 position, so the meaningful question is whether they improve on owning the index — hence the S&P 500 benchmark. Strategy 4 is a standalone, broadly market-neutral intraday premium program with no core index exposure; benchmarking it against the S&P 500 would compare unlike things. The relevant comparison is the naive alternative that shares its mechanics — selling a 0DTE straddle every session — which isolates the only thing the model claims to add: selecting when to trade and when to stay flat. Outperforming that benchmark does not imply low absolute risk, since the benchmark is itself a high-risk approach.

Greeks / execution. Greeks monitored, not targeted. Execution per Common Framework, with the added caution that same-day settlement and extreme 0DTE gamma make entry timing and fill quality especially consequential.

Backtest methodology

(Sample period: 2024-02-08 – 2026-06-30; hypothetical)

Average outperformance vs benchmark (0DTE straddle sold every session): + 0.26 % per month, + 0.83 % per quarter, + 2.79 % per year
Average time in trade (holding duration): 1 day