A 0DTE strangle — one cheap out-of-the-money call, one cheap out-of-the-money put, bought together before the market picks a direction — is one of the most attractive-looking ideas in options. The risk is capped at a few dollars, the payoff is convex, and you do not have to be right about direction. On the day it works it is spectacular.

We had one of those days on 3 September 2026, with real fills. Then we tested the same structure on 374 sessions of real traded option prints, and the two answers are so far apart that the gap is the lesson.

ℹ️ INFO
**Where this came from.** This began as an internal idea we named **Oyamori TwinStrike** — arm both tails cheaply during a consolidation, harvest whichever leg expands, keep the other as a free reversal ticket. This article is the result of trying to verify it. The name stays with us for a specification that clears the bar; what is tested below is the specific version described here, and the verdict applies to that version.

The day it works, priced from the actual chain

Here is the real 0DTE chain at 10:16 ET on 3 September, SPY at 768.61:

SPY 0DTE option chain at 10:16 ET showing the 773 call at 0.06/0.07 and the 764 put at 0.15/0.16
The real quotes: 773 call bid 0.06 / ask 0.07, 764 put bid 0.15 / ask 0.16. A $23 campaign, both legs armed.

SPY then ran to a session high of 773.37 by 14:15. Repricing the call along the actual path — calibrated to the real $0.07 fill rather than to any displayed implied volatility — gives:

$23
Campaign cost
$0.56
Call at 11:35, SPY 772.00
$0.82
Call at 14:15, SPY 773.37
+$59
Campaign profit at that exit
+257%
Return on debit

That is a real trade with real numbers, and it is genuinely excellent. It is also why this idea is so hard to let go of.

⚠️ WARNING
**One calibration trap worth knowing before you model any of this yourself.** The broker displayed an implied volatility of 13.64% for that 773 call. Price the contract with that number and you get **$0.88**. The market's actual ask was **$0.07** — the displayed figure overprices a far-OTM 0DTE ticket by roughly **13×**. Any backtest of cheap options that models premiums from a displayed IV field is producing fiction. Everything below uses traded prints instead.

The problem: cheap enough and close enough are different strikes

The strategy has two rules that sound compatible and are not: buy 4–5 points out of the money, and keep each leg between $0.05 and $0.20.

We measured what those distances actually cost. Real traded 0DTE prints at 10:15 ET across 65 SPY sessions:

Distance Call Put Campaign Either strike touched within 60 min
2 pts $1.47 $0.46 $193 49%
3 pts $1.02 $0.34 $136 26%
4 pts $0.64 $0.27 $91 14%
5 pts $0.40 $0.20 $60 8%
6 pts $0.25 $0.15 $40 6%
7 pts $0.14 $0.12 $26 2%
8 pts $0.09 $0.10 $19 2%
9 pts $0.06 $0.08 $14 0%

At 4 points the campaign costs $91 — three times the budget. At 7 points it costs $26 and the underlying gets there 2% of the time. On the measured medians, 86% of sessions blow a $30 cap if you insist on 4–5 points on both sides.

Chart showing campaign cost falling and touch rate falling together with distance, with no overlap between affordable and reachable
Cost and reach fall together. The bars inside the green band are the affordable ones; the amber line is how often price actually arrives. They do not overlap.

Move the slider and watch the two conditions refuse to meet:

⛶ Full Screen

🚨 DANGER
**And "buy the $0.12 contract" is not a fixed rule.** The same nominal 4–5 point call traded at **$0.18** on the cheapest day in twenty and **$1.87** on the richest — a **10× range**. A premium-anchored rule therefore buys a *different distance every day*: further out when the market is quiet, closer when it is volatile. That is a hidden bet on volatility regime wearing the costume of a fixed rule.

The mechanism does not hold either

The entry logic says: wait for a mature compression, because compression precedes expansion. That is testable directly, without any options involved.

Measured on SPY 1-minute bars, day-clustered so each session counts once:

+0.0055 ATR
Paired difference
−0.0071 to +0.0182
95% interval
236
Sessions
null
Verdict

Compression does not predict expansion. Low volatility mostly begets more low volatility.

The trap we nearly published instead

The first version of this test sampled every minute with a 60-minute forward window and returned what looked like a decisive answer at n = 103,014. That n was fictional — overlapping windows on the same session are not independent observations, and the real sample was 236 sessions. Re-run day-clustered, the same data gives a clean null. Any study of intraday setups that reports five- or six-figure sample sizes from overlapping windows is reporting a confidence interval that cannot be true.

There is a second-order consequence that matters more than the null itself: the compression rule tends to arm around 10:48 ET — into the quietest, most theta-heavy stretch of the day. The trigger does not just fail to help; it systematically selects the worst hour to own a decaying option.

What the structure actually returned

On real traded 1-minute option bars, 374 sessions, both fill conventions:

Configuration n Return on debit 95% interval
Blind arm 09:45, harvest 1.5×, optimistic fill 283 −19.8% ±8.9
Blind arm 09:45, harvest 2.0×, pessimistic fill 283 −35.3% ±10.6
Compression-armed, harvest 2.0×, optimistic 189 −44.9% ±13.0
Compression-armed, harvest 3.0×, optimistic 189 −59.5% ±14.2
Compression-armed, 2.0×, 60-minute stop 189 −10.0% ±6.0

Every cell negative. Every interval excludes zero. And the trigger made it worse — −44.9% armed against −26.2% blind on the same strikes, ladder and fills.

⛶ Full Screen

⚠️ WARNING
**The single biggest source of error in this kind of test is not the strategy, it is the fill assumption.** In this run, one cell came out at **+4.1%** using the closing print of the arming minute as the entry, and **−25.4%** using the worst print of that same minute. Same trade, same data, one convention — a 30-point swing. Any published result on a structure this cheap that quotes one convention and not the other has not answered the question.

The second strike almost never arrives

The most appealing part of the idea is the leg you keep: the loser rides on as a free ticket in case the move reverses.

Of the sessions where the first leg reached 2× the campaign cost, the opposite leg subsequently reached 2× in 0 of 9, with a median subsequent peak of 0.22× campaign cost.

The reason is mechanical rather than bad luck. Once one side expands, the other is further out of the money and has less time remaining — both inputs move against it simultaneously. And it was never free: you paid for it in the debit. Calling it a free ticket is mental accounting, and on 3 September that "free" put was 70% of the entire campaign cost for a leg that was never once live.

Even with perfect foresight and zero spread — exiting at the single highest print of the session — the premium-anchored version reaches 2× campaign cost on only 25% of sessions.

The part that generalises: you can refute this, you cannot confirm it

This is the finding we would keep if we had to throw away everything else.

The measured effect here is a loss of 10–60% of debit. The minimum detectable effect at the sample sizes available is:

Campaigns MDE, hold to expiry MDE, 60-min stop
100 ±25.5 pp ±11.8 pp
200 ±18.0 pp ±8.3 pp
400 ±12.7 pp ±5.9 pp

Now count what SPY can actually supply. About 630 sessions have traded 0DTE option bars; a compression rule arms on 74% of them; the target contract has a print at the arming minute on 76% of those; split off a holdout and you are left with roughly 100 holdout campaigns.

🚨 DANGER
**At n ≈ 100, the interval is ±12 to ±26 percentage points of debit.** A genuine 10%-per-campaign edge is invisible at that resolution. So this study was **capable of refuting and structurally incapable of confirming** — and so is anyone else's. When you see "my 0DTE strangle strategy works, here's a year of results," the arithmetic says they have not measured that, whatever the equity curve looks like.

The asymmetry is the useful part: a large negative is detectable at n = 100, a modest positive is not.

What we are committing to, in public

Because a scoping run cannot confirm anything, here is the specification we will run, declared before we look — so that if it comes back positive it means something:

ARM      first minute where the trailing 20-bar mean range ÷ 14-day ATR is at or below
         the 30th percentile of the PRIOR 20 sessions, between 09:50 and 14:30 ET.
         One campaign per session. No-arm sessions recorded, not discarded.
ARMS     A: strike nearest 0.55 ATR out of the money, no premium constraint
         B: strike nearest $0.12 within the $0.05–0.20 band, distance unconstrained
LADDER   harvest the first leg at 1.5x campaign cost; secondary rungs 2.0x, 3.0x
EXITS    primary 60-minute mark-to-last-print; secondary hold to expiry
FILLS    reported as a bracket — optimistic and pessimistic — never one alone
METRIC   return on debit per campaign, day-clustered bootstrap interval,
         reported with median, P(r>0), the top-3 share of total profit, and log growth
CONTROLS random minute, the trigger's own off-state, single-leg, time-matched random
HOLDOUT  untouched from 2026-03-01 onward; training figure committed in writing first
GUARD    no point estimate published below 400 campaigns

And the four things that would change our minds, also declared now: a pessimistic-fill result positive with an interval excluding zero on the holdout; the effect surviving both the random-minute and off-state controls; the compression mechanism replicating on its own; and positive log growth, not just a positive average.

Why a refutation can be published from a scoping run when a confirmation cannot

Pre-registration exists to stop an analyst fishing for a positive result — trying rules until one works, then presenting it as though it were the first thing tried. That process cannot manufacture a negative. If anything the bias runs the other way: an analyst picks the rule they believe will work, so a rule chosen freely and still losing 20–60% of debit across every cell is evidence pointing away from the hypothesis more safely than the same process pointing toward it. That is why the grid above is publishable as directional evidence, and why the pre-registered spec still has to be run before anyone — including us — claims the opposite.

So what do you do with cheap convexity?

Not nothing. Three honest uses survive everything above:

Still defensibleNot supported
Event riskbuying a tail before a known binary, sized as insurance—
Sizinga defined-loss ticket as a small share of a larger plan—
Readingthe chain tells you what the market thinks is reachable—
Costpay for reach, not for cheapness—
As a daily strategy—−10% to −60% of debit per campaign, 374 sessions
Compression as the trigger—null mechanism, and it arms at the worst hour
The free reversal ticket—0 of 9, median subsequent peak 0.22x
Cheapness as the strike rule—it buys distance you cannot reach

The deepest problem is that cheapness and reachability are the same axis viewed from opposite ends. A $0.07 ticket is $0.07 because the market has priced how rarely it pays — and the market is not wrong about that in any way we have been able to measure. This is the same conclusion our 0DTE strike selection work reached from the reachability side, and the same one the cost floor reaches from the spread side.

The takeaway

  1. Price the campaign before you admire the payoff. At 4 points out, both legs cost about $91, not $23. The $23 version lives 7 points out, where price arrives 2% of the time.
  2. Never model a cheap option's premium from a displayed IV. It overpriced our real $0.07 ticket by 13×.
  3. Compression does not predict expansion — +0.0055 ATR, interval [−0.0071, +0.0182], n = 236 sessions.
  4. The kept leg is not free. You paid for it, and it reached 2× in 0 of 9 opportunities.
  5. Ask what n your evidence has. Below a few hundred campaigns you can detect a disaster and not an edge — which means a good-looking year proves far less than it feels like it does.

The day it works will always be more vivid than the 373 that did not. That asymmetry of memory is the actual opponent.

The one line to keep
A cheap option is cheap because of how rarely it pays. You can buy reach or you can buy cheapness — the market does not sell both at once.
ℹ️ INFO
**Method.** Real traded 1-minute SPY 0DTE option bars (Alpaca) and matched 1-minute underlying paths. Premium-vs-distance: 65 sessions, Mar–Aug 2026, median traded print at 10:15 ET. Campaign outcomes: 374 sessions, Mar 2025 – Aug 2026, n = 172–283 per cell — **a scoping run, below our own n ≥ 400 guard, with the compression rule and ladder chosen before results were seen but not pre-registered.** Sessions where the target contract had no print at the arming minute were dropped (24% distance-anchored, 46% premium-anchored), a liquidity selection of unmeasured sign. Intervals are normal-approximation and day-independent; a clustered bootstrap would be wider. One volatility regime. Nothing here is investment advice, and the verdict applies to the specification described, not to every possible version of the idea.