Options for Quants · Systems Research

Anatomy of a Straddle Engine

A long ATM straddle is the cleanest bet in options — pay theta, get paid by movement in either direction. Running one systematically is another matter: when is movement actually coming, when is the premium too rich to pay, and how do you stop a +2% peak from round-tripping into a stop-loss? This note opens up our production engine — the entry vote, the veto gates, the single profit governor — and ends with an event study on 18 months of 5-minute data that shows exactly what the entry signal does and does not predict.

By Options for Quants Venues NIFTY · BANKNIFTY · SENSEX · FINNIFTY · MIDCPNIFTY Data 5-min bars, Mar 2025 → Sep 2026 Study 953 signal events, path-level

1 · The entry decision: one hard gate, a vote, seven vetoes

trigger → quality → price-of-vol

The engine never enters on opinion. Entry requires three layers to align, evaluated every scan cycle:

Layer 1 · Mandatory event

CUSUM ≠ 0

A signed price-CUSUM (calibrated drift threshold, reset each session) must have fired. No event, no entry — full stop. The rule is empirical: every early stop-loss in our live record entered without a trigger. The event latches for several scan cycles so the downstream checks can complete without the trigger expiring mid-entry.

Layer 2 · Quality vote

≥ 3 of 4

Four booleans: the CUSUM event · vol-squeeze momentum positive (or a vol-CUSUM breakout, or strictly rising 3 bars) · the Kalman vol filter's own uncertainty above a high percentile of its own history · vol-of-vol elevated versus its median. Since the event is already mandatory, this means the trigger plus at least two vol-quality votes: the event says something is moving; the votes say the vol surface looks pre-expansion, not post.

Layer 3 · Veto gates

7 vetoes

Passing the vote buys nothing yet. Each gate below can still refuse — and each refusal is logged with its reason, because a silent gate is an unauditable gate.

Veto	Rule	Why it exists
VOL CONTRACTION	VSM < 0 + uncertainty low + no breakout → block	Entering into contraction triggers an immediate thesis-death exit — you'd pay two spreads for nothing. Overridden when the CUSUM just fired or VSM is inflecting up: momentum outruns lagging vol metrics.
VRP	IV − RV above a calibrated cap → block	The straddle is priced richer than what the market has actually been delivering. Buying expensive vol is how theta wins.
RV / IV	RV_adj / IV above a calibrated floor > 1	The multiplicative version of the same idea — and RV must be overnight-adjusted (√(intraday² + gap²)). Intraday Parkinson vol misses close-to-open gaps, roughly half of index variance; on this gate's first live day the unadjusted ratio wrongly blocked every single evaluation.
IV SKEW	\|IV_CE − IV_PE\| / mean above a calibrated cap → block	Materially different leg IVs mean the market is pricing directional drift — the rich leg bleeds faster than the cheap leg gains. That's not a pure vol bet any more.
IV WARMUP	vol-regime model not yet warmed → block	Without regime context you can't know if vol is cheap or expensive. Don't enter blind.
IV VELOCITY	dIV/dt falling → block	Long premium into declining IV starts the trade with a vega headwind.
TIMING	no entry 09:15–09:30 · expiry day after noon · DTE < 1	The open IV spike marks entries against you; sub-1-DTE gamma is a different (and crueler) game.

Only after all of that does the order path run — and it has its own gauntlet: a depth check (both legs must show live bid and ask), a five-level order-book walk on thin symbols, a cost gate (round-trip cost must stay under a fixed share of the premium at risk), and a margin precheck. Every refusal is a retryable abort: the scanner simply re-fires next cycle.

2 · The exit problem is the real problem

one governor, not a committee

Our trade history taught us that the straddle's P&L problem was never entries — it was giving gains back. Worse, the natural evolution of an exit stack (a trailing stop, a profit target, a z-based ride) creates rules that fight each other: the tight trail exits the trade the ride wanted to hold. Two design decisions fixed this:

A single gain governor. Once net P&L crosses an objective anchor, one rule owns the profit side: if the displacement z-score is still expanded, ride with a rollover trigger and a gain floor; if z has collapsed, bank now. While the governor is armed the legacy trailing stack is suspended — two giveback ratchets can never fight.
A sub-target giveback floor. The governor created a dead band: peaks below the anchor had no profit exit at all, and live specimens round-tripped +2% peaks into full stops. The floor arms once the peak clears a calibrated bar, gives back a fixed fraction of the peak, and requires the breach to persist for several consecutive cycles — one wide-bid flicker of a quote cannot fire it. It was validated by replaying every historical trade's per-cycle P&L path before it touched production.

Exit	Fires when
ADAPTIVE STOP	a hard floor, widened in proportion to the entry IV's implied daily move — high-IV entries earn wider stops; big swings were priced in
DELTA CAP	\|\|Δ_CE\| − \|Δ_PE\|\| beyond a calibrated cap — the straddle has drifted into a directional bet; it no longer is the trade you put on
GAMMA LOCK	\|position Δ\| over cap while profitable — lock the rally before it mean-reverts
GOVERNOR	objective anchor reached → bank or ride by z-state (the only profit-taker while armed)
GIVEBACK FLOOR	once the peak clears an arming bar: a fraction-of-peak floor, multi-cycle breach confirm
VOL CONTRACTION	squeeze momentum collapses — the expansion we bought never came; stop paying theta for it
STALE / MAX TIME	no progress — but vetoed if the CUSUM is near a fresh trigger or the Kalman innovation just spiked: never pay exit costs seconds before the move
BOUNDARIES	expiry-day and end-of-day squareoffs that no profit state can suppress

3 · What does the entry signal actually predict?

event study · 953 signals · path-level

Here is the honest part. We recently tested a stronger trigger — the CUSUM event confirmed by a Kalman-filtered trend z-score, combined into a confidence score with a 30-minute latch — as a candidate directional signal. On fixed 30-minute horizons it looked mildly directional (54% hit rate at high confidence). Then we followed every one of the 953 confirmed signals bar-by-bar for two hours:

Three panels: flat mean signed path inside a wide inter-quartile band; symmetric favorable vs adverse excursion scatter; monotone rise in excursion magnitude from random control to Kalman-confirmed signals while favorable-first stays at 50% — **The signed path is flat** (peak drift +0.04% inside a ±0.25% inter-quartile band); **favorable and adverse excursions are mirror images** (medians +0.25% / −0.24%); and the first-passage test — which ±0.25% barrier gets hit first — comes out 51%, a coin flip. The fixed-horizon 54% did not survive path-based measurement. But panel (c) shows the effect that is real: **median two-hour excursion magnitude rises monotonically** from 0.14% at random bars to 0.19% after a raw CUSUM, 0.24% after the confirmed signal, 0.27% after a strong Kalman z — while directional preference stays at 50% on every rung.

BANKNIFTY, 15 sessions of 5-minute closes: 50 raw CUSUM triangles reduced to 7 circled Kalman-confirmed signals — The confirmation filter at work: 15 sessions of BANKNIFTY, 50 raw CUSUM events (triangles) reduced to 7 Kalman-confirmed signals (circles). One caution from the same chart: the lone spike at a session close is a bad settlement print that manufactured a phantom signal — close-bar events are excluded from our journals, and every data pipeline should assume its final bar of the day is guilty until proven innocent.

Verdict: these signals are movement detectors, not direction detectors. After a confirmed event, price travels ~70% further than baseline within two hours — with no preference whatsoever for the signal's direction. That is a finding with teeth: it says the trigger belongs in front of a long straddle (bounded loss, paid by movement, indifferent to direction), and emphatically not in front of a debit vertical. It is also the third independent time our data has produced this exact shape — signed information in vol-triggered events is noise; unsigned information is activity.

Candidate triggers like this one never go straight to production here. They run in a premium-marked shadow first: every signal journals the option structure it would have bought at live bid/ask, marked again at 30, 60 and 120 minutes — because a spot-move edge can die completely in the option premium (spreads, theta, and the fixed cost of four legs). Arming waits for the shadow, not the backtest.

Method & limitations

Setup. 5-minute bars, five index underlyings, Mar 2025 – Sep 2026 (~363 sessions each). CUSUM: signed price-drift detector with session reset (production calibration withheld). Kalman: local-level price filter; the confirmation score blends trigger freshness (full weight ≤ 30 min, decaying to zero by 75) with |z| agreement, and a confirmed verdict latches 30 minutes against noise. Event study measures signed forward paths at signal starts with same-session windows only; first-passage = which of ±b barriers is touched first, b ∈ {0.15%, 0.25%, 0.40%}.

Spot moves are not option P&L. The movement effect is measured on the underlying; whether it survives straddle premium, spread and theta is exactly what the live premium-marked shadow exists to answer.
One-and-a-half years, one market. Indian index options, a window containing both rising and falling vol — but no guarantee of stationarity.
Calibrations are withheld. The exact gate values were tuned on our fills and our cost structure and are part of the system's edge; the architecture is the template. The experiment specification for the event study, by contrast, is given in full.
Survivor-free accounting. The 953 events include every confirmed signal, not just the ones followed by trades.

Companion note: The Volatility Clock — intraday IV behaviour across the trading day.

Disclaimer. This is a research note for informational and educational purposes only and does not constitute investment advice or a recommendation to buy or sell any security. Options and F&O trading involves significant risk of loss. System parameters and study results are drawn from a limited historical window and our own execution environment; past behaviour is not indicative of future results. Spot-level findings do not account for transaction costs, slippage, or liquidity. Please read all scheme-related documents carefully and consult a qualified adviser before trading.