Opening Range Program

01The strategy

One bet per session, fully systematic, no discretion

At 09:35 New York the strategy reads the first five-minute bar of the regular session. If its close is above the 12-period EMA of prior five-minute closes, it buys; below, it sells short. The stop sits at 2× the true high–low range of that opening bar, and position size is set so that the distance from entry to stop is exactly 1% of equity. Leverage is capped at 2×. Any position still open at 15:30 is flattened into the close. Nothing is held overnight.

That is the entire specification. There are no filters, no regime switches, no discretionary overrides, and no parameters beyond the stop multiple and the instrument — both of which are examined in §3.

02Performance record

Year by year, net of costs and measured slippage

† partial year. Returns compound within the year; sizing resets to 1% risk per trade every session.

Overlapping one-year windows

Calendar years are an arbitrary slice. Across all 2258 overlapping 252-session windows in the sample, the Sharpe ranged from 0.06 to 2.36 at the 95th percentile, with a median of 1.37. One-year returns ranged from -0.8% to +65.4%, median +25.7%. The fraction of one-year windows that lost money was 1 of 2258, and the worst of them lost -0.8%.

03Parameter surface

Is the edge a peak, or a plateau?

A strategy whose performance sits on a spike in parameter space has been fitted to its sample. One whose neighbours all work has found something. The two grids below are the test. Every cell is an independent backtest over the full sample at the stated configuration.

QQQ is the strongest of these five at every stop multiple, but it is not alone — SPY carries the same signature at lower amplitude, which is what a real intraday-momentum effect should look like across correlated equity indices. TLT and GLD, different asset classes with different opening dynamics, do not. Five instruments is too few to settle anything; §4 runs the same frozen rule across 95.

Across calendar years the surface stays positive nearly everywhere, and the ordering of stop multiples is stable. This is the visual counterpart to the combinatorial purged cross-validation reported in the research file: 96% of 455 training splits independently selected QQQ at 2×, and out-of-sample path mean Sharpe came in at +1.53 against an in-sample best of +1.57 — a selection bias of +0.03.

04Breadth: does this exist anywhere else?

95 instruments, one frozen rule, every result reported

Everything to this point describes one instrument, and that cuts two ways. It is a statistical problem: a single instrument is a single test, and from inside it there is no way to tell whether the effect is a property of markets or a property of this ticker. It is also an allocation problem: a programme that works in exactly one place cannot be scaled by adding places, so the capacity ceiling in §8 becomes the whole business rather than one sleeve of it.

So the identical rule — same EMA-12, same true high–low opening range, same 2× stop, same 15:30 flatten, same 1% risk sizing — was run across every instrument reachable from the data subscription: 95 instruments spanning US index funds, all eleven sector SPDRs, industry funds, single stocks, single-country and regional international funds, government and credit bonds, commodities, currencies, volatility, crypto spot and crypto ETFs. All 95 returned enough history to test, so nothing was dropped for a reason that could be confused with a result.

Two data problems that had to be fixed first

The deployed signal seeds its EMA from an hour of pre-market five-minute bars. QQQ has that on essentially every session. XLK has it on 64% of them, and thinner instruments far less. Requiring it would silently discard two thirds of the sample across most of this universe, and would keep precisely the sessions with unusual pre-market activity — a selection, not a sample. So the EMA is computed over the last twelve five-minute closes on the continuous tape, reaching into the prior session where pre-market is thin. On the 36 instruments with near-complete pre-market coverage — where the deployed definition is the one actually running — the two produce the same side on a median of 99.8% of sessions and never below 95.5%. Across all 74 instruments where both are defined the median is 98.9%. That is what licenses the substitution.

The second problem was cost. The original plan used the Corwin–Schultz (2012) high–low spread estimator, which needs no quote data and so runs everywhere for free. It failed calibration and was discarded:

Basis points. The estimator identifies the spread as the part of a two-day high–low range that does not scale with the square root of time. On a modern liquid fund the spread is a rounding error next to the volatility, so that identification is noise — over half of QQQ's daily estimates come out negative. Truncating symmetric noise at zero, as the method prescribes, manufactures the bias. It is a sound estimator for markets whose spreads are wide relative to volatility — and the table shows exactly that: the overstatement falls from 59× on QQQ to 7× on UNG as the real spread widens. Nothing in this universe is wide enough. It is shown rather than quietly dropped, because a rejected method with its calibration attached is more useful than a method that never appears.

Spreads are therefore measured, on the same consolidated tape and through the same endpoint as the slippage work in §5, at 09:35 and at 15:30, on a common grid of 120 sessions spanning the full window. The round-trip charge is half the entry spread plus half the exit spread. Median across the universe: 2.57bp, against the 0.5bp the QQQ backtest assumes — and QQQ's own measured figure is 0.53bp, which is the first thing this exercise established: the deployed cost assumption is right for the instrument it was written for, and wrong for almost everything else.

The cross-section

The grid below charges every instrument the same flat 0.5bp. That is deliberate and it is not the answer — it isolates the question of where markets behave this way, holding cost constant so that the only thing varying between rows is the market itself.

Grouped by the exposure an allocator would actually be buying, with the measured cost applied in the final column:

"Positive" and "median Sharpe" are at the flat 0.5bp. Spreads in basis points, measured. Leveraged funds and crypto ETFs are shown but excluded from every significance count, because a 3× Nasdaq fund is the same bet as QQQ rather than independent evidence about it.

What the coin flip says

The single most important control on this page. The strategy's payoff is asymmetric by construction — the stop truncates the loss, the flatten does not truncate the gain — so a coin flip run through the identical machinery is not centred on zero, and any test that assumes it is will read structure as skill. The null is therefore generated, twice: once at the flat cost and once at each instrument's own.

Both readings matter and they say different things.

The effect is real and it is broad. At a constant cost the rule returns a mean Sharpe of 0.17 across 85 independent instruments where a coin flip returns -0.01 and reaches only 0.05 at the 95th percentile of 2,000 draws — p = 0.0005. The direction call carries information that the stop structure alone does not, and it carries it in markets the rule was never tuned on.

The effect is mostly not worth what it costs to harvest. Charge each instrument its own measured spread and the same cross-section turns over: mean Sharpe -0.60, with only 26 of 85 instruments still positive. It is still significant — a coin flip charged the same spreads returns -0.77 and gets only 15% of instruments positive, p = 0.0005 — but "significantly better than losing money faster" is not an investment case. In most of this universe the signal is real and the spread is bigger.

Multiplicity

With 85 independent instruments tested, some will look significant by chance. Benjamini–Hochberg at a 10% false-discovery rate returns 8 discoveries at the flat cost and 8 at measured cost; the far stricter Bonferroni bound at α = 0.05 returns 2 and 3.

Both corrections assume independent tests and these are not independent. The mean pairwise correlation of strategy returns is 0.08 — low, because most of these instruments contribute noise rather than a common signal — which under an equicorrelation approximation still makes 85 instruments worth roughly 11 independent tests rather than 85. Bonferroni against 85 therefore over-corrects. The honest reading sits between the two figures, and it is stated that way rather than resolved, because resolving it would mean choosing the correction that produces the preferred answer.

Every instrument, individually

Ranked by what survives its own spread. p is the permutation p-value against that instrument's own coin-flip null at the matching cost; a dot marks p < 0.05 at measured cost. Breakeven slip is the fraction of the stop distance at which the Sharpe crosses zero — measured slippage on QQQ is 0.0079. Beta is against the instrument's own regular session. Median daily volume is measured over 09:30–16:00 only, and below $100m it is marked in red: those instruments cannot absorb size regardless of what their Sharpe says. The crypto rows are the exception to read carefully — their volume comes from one venue's feed rather than a consolidated tape, so it understates the real market by a wide margin and should not be compared with the equity rows. "dep." marks instruments excluded from significance counts as duplicates of an underlying already in the table.

What the surviving instruments have in common

The instruments that clear their own costs are not scattered. Leading the table are QQQ itself, then MSFT, XLK and SMH — the Nasdaq-100 and its largest constituent, the technology sector fund, and the semiconductor fund. SPY follows. The one non-US name near the top is EWY, which is a Korea fund in name and a semiconductor fund in composition. The effect concentrates in large-cap technology and semiconductor index exposure, and decays smoothly as instruments move away from that centre: IWM 0.20, DIA -0.34, and the defensive sectors below zero.

One entry near the top of that table is there to be distrusted. FXB, a sterling fund, posts the highest flat-cost Sharpe in the whole universe at 1.81, and still clears its own measured spread at 0.61 with p = 0.004. It trades $4m a day. It passes every statistical test on this page and could not absorb a single session of the programme at its current size — which is why median volume is a column in that table rather than a footnote. Statistical significance and investability are different properties, and this universe contains instruments with one and not the other.

The concentration pattern was not designed for and it is a post-hoc observation, so it is offered as a description rather than a mechanism. What supports it as more than a coincidence is the leveraged row. TQQQ, QLD and SQQQ — the 3× inverse fund — all reproduce the result at 1.34, 1.21 and 1.04. A rule that works on both a fund and its inverse is not making a directional bet on the index; it is reading something about how that index trades in its first five minutes.

Futures, and why crypto is the interesting null

There is no futures endpoint on this data subscription, so the equity index, metals, energy, rates and FX complexes are represented by their ETFs. Those proxies carry the same underlying exposure but not the same execution economics — futures quote tighter in percentage terms and carry no borrow — so an ETF result should be read as a floor rather than an estimate. Given how decisive costs turn out to be, that gap is material and it is the single most valuable extension of this work.

There is also a mechanical reason to expect futures to behave differently, and it belongs here as a prediction rather than an excuse. The rule trades the resolution of an overnight order imbalance into a single opening auction. A contract that trades through the night never accumulates one. Crypto is the clean test, because it trades continuously and 09:30 New York is not an open at all. At a flat cost it produces ETH/USD 0.60, DOGE/USD 0.32, BTC/USD 0.24, SOL/USD -0.40, LTC/USD -0.95 — noise. At measured spreads of BTC 14, ETH 14, SOL 29, LTC 44, DOGE 44bp it produces a uniform disaster. Both the signal and the economics fail exactly where the mechanism says they should.

What a diversified version would actually be

The allocator's question is whether this can be sold as multi-market exposure. The construction below takes the most liquid independent instrument in each asset class — chosen on median dollar volume, which is knowable in advance, never on the result — and combines the legs at equal risk, scaled to a common daily volatility target. Every leg runs the same frozen rule.

That last line is the one worth sitting with. A diversified sleeve should beat its best single component on a risk-adjusted basis; that is the entire argument for diversification. Here it does not, even when the components are hand-picked with hindsight, because the legs are close to uncorrelated in their noise and share almost none of their signal. Adding them dilutes the one instrument that works.

What this section does and does not establish about the other 94

QQQ is reportable because four separate things have been done to it. This section does the first of them for 94 more instruments, and only the first:

• A parameter selection validated by combinatorial purged cross-validation, with the selection bias measured rather than assumed. Not done elsewhere — the frozen-config test here is a different and weaker claim, deliberately so.
• Stop fills measured against that instrument's own consolidated tape. Not done elsewhere. The measured QQQ slippage of 0.0079 is applied to every instrument in this section, which is optimistic for all of them: a wider-quoting instrument slips further past its stop, and its own breakeven column is correspondingly closer. Entry and exit spreads are measured per instrument; stop-fill slippage is not.
• A forward distribution from block-bootstrapped paths, including the halved-edge case. Not done elsewhere.
• A capacity statement from that instrument's own opening-window turnover and displayed depth. Not done elsewhere — median dollar volume is reported as a screen, which is not the same thing.

So nothing in this section makes a second instrument deployable. It answers a narrower question — whether the effect is a property of markets or of one ticker's ten-year sample — and that is all it should be read as answering.

So the honest answer to the breadth question has two halves, and neither should be quoted without the other. The effect is not an artefact of fitting one ticker: it appears across markets the rule was never tuned on, at p = 0.0005 against a properly constructed null, with a beta to the underlying market of essentially zero — so it is neither curve-fitted nor disguised long exposure. And it is not a multi-market programme: once each market is charged what it costs to trade, the investable set collapses to large-cap US technology index exposure and a short list of things that behave like it. This is a real, narrow, capacity-limited effect. §8 puts a number on how narrow.

05Execution: what a stop actually costs

The measurement that decides whether any of the above is real

Position size is derived from the stop distance, so the entire risk model depends on stops filling near their price. The stop is hit on 40.8% of sessions. If a stop-market fill gives up a large fraction of the stop distance, the strategy pays that toll four sessions in ten and the edge disappears.

An earlier internal study put that cost at 28–41% of the stop distance against a breakeven of roughly 29%, and the programme was shelved on that basis. That study was wrong — not conservative, wrong. It located the first minute close beyond the stop and called the gap slippage, which measures up to 59 seconds of continued adverse price movement. That movement is the market going against the position, which the strategy's returns already pay for. The number double-counted a loss the backtest was already taking.

The replacement measurement

Every stop event is now replayed against the consolidated tape:

1. Locate the five-minute bar whose high or low first pierces the stop.
2. Pull every SIP print in that bar and find the first trade at or through the stop price. That print is the trigger instant, to the nanosecond.
3. Take the national best bid and offer at trigger + 50ms of order latency.
4. Fill on the far side — a long's sell-stop hits the bid, a short's buy-stop lifts the offer — and measure the adverse distance from the stop price as a fraction of the stop distance.

Sample: 497 stop events drawn at random across the full ten years.

By year

The measured cost is roughly two orders of magnitude below the figure that shelved the programme, and it is stable across the decade including 2018, 2020 and 2022. The mechanism is unremarkable: QQQ quotes a one-cent spread against a stop distance of three to seven dollars, and the order sizes involved rest inside the displayed top of book.

Phantom stops — a finding about the backtest, not the fills

Building this measurement surfaced a separate defect. A five-minute bar's high and low include every print, odd lots and errors alike, and the backtest triggers stops on those extremes. On 2026-01-15 the 10:55 bar carried a high of exactly 630.00 from two prints of 40 and 20 shares on one venue, two microseconds apart, while the NBBO sat at 625.29 / 625.37. The backtest books a stop-out there. The order book never offered one.

Every trigger is therefore now classified before it is measured, and only round-lot, regular-way prints corroborated by an NBBO that also reached the stop enter the distribution above. Of 500 sampled stop-outs, 3 (0.6%) failed that test. The direction matters: a phantom stop closes a trade that was still running, so it removes the right tail. The backtest is pessimistic on those sessions, and the headline figures in this document have not been adjusted upward for it.

This grid is the whole argument. Read down the left column: at a slippage of 0.29 and above, every configuration is dead, which is exactly why the programme was shelved. The measured value sits in the top row. The strategy was never fill-limited; the instrument used to measure the fills was broken.

06Forward expectations

Bootstrap distributions, not a point estimate

Past returns are one path. To say anything about future ones, the daily return series is resampled in 21-day blocks — preserving volatility clustering and short-horizon autocorrelation that an independent resample would destroy — into 10,000 synthetic paths per horizon.

The second block halves the mean return while leaving volatility untouched. That is the honest way to stress an edge: strategies decay by earning less, not by becoming safer.

Percentiles across 10,000 bootstrap paths. Annualised figures are compounded from terminal wealth. Drawdown is measured within each path.

The lower half of the table is the number to underwrite against. A programme that still clears its drawdown budget with half its edge removed is one whose risk model does not depend on the edge being exactly as measured.

07Return concentration

How much of the edge lives in how few sessions

This is the least comfortable table in the document and it is presented deliberately. The return profile is positive-skew: a hit rate near 45.3% with winners materially larger than losers, which means a small number of sessions carry a large share of the result. Skew of daily returns is 1.19.

The mitigation is frequency, not diversification: with roughly ten such sessions per year, the annual result is reliable even though the daily one is not — which is what the year table in §2 shows. The exposure this creates is operational. Downtime is not an inconvenience; a missed session is a lottery ticket not bought. Uptime is a first-order risk control for this programme and is treated as one.

08Capacity

The binding constraint, stated plainly

Sizing to 1% risk against a stop 0.46–0.99% wide puts notional exposure at 1.0–2.2× equity, established inside one five-minute window and unwound inside another. Measured over the six months to August 2026, QQQ turns over a median of $29.8bn a day, of which $2.1bn prints in the first ten minutes — about 5.9% of the session.

Holding participation to 10% of that opening window caps notional near $210m, which puts equity capacity in the region of $100–200m before market impact starts eating the edge the slippage measurement just established. Two things move that number: executing in Nasdaq-100 futures rather than the ETF, which is a materially deeper book, and spreading the exit across a longer window, which trades impact for timing risk against the 15:30 flatten.

Displayed depth is the reason to treat that ceiling as real rather than conservative. Across the 497 measured stop events, the size resting on the exit side of the NBBO was a median of 600 shares and 100 shares at the tenth percentile — on the order of $200k at current prices. The measured slippage is therefore valid at the sizes tested and says nothing about an order that has to walk the book.

This is the honest ceiling, and it is the reason this document describes a sleeve, not a fund. The profile — Sharpe near 1.5, Calmar near 2, worst underwater period under a year — is the shape of a well-behaved book within a multi-strategy platform. It is not the shape of a standalone product, because at this capacity the fee stream does not support one.

09Risks and limitations

Stated without softening

• No live track record. Every figure in this document is a backtest. The paper account has traded a handful of sessions — far too few to carry any statistical weight, in either direction.
• Statistical power. Distinguishing a true Sharpe of 1.5 from zero using live returns alone requires roughly 1.7 years of trading for a t-statistic of 2, and about 2.6 years for a properly powered one-sided test. After a single year, a Sharpe estimate carries a standard error near 1.0. Live P&L is the slowest available instrument, which is why execution is validated by direct measurement instead.
• One instrument, one bet per session. There is no internal diversification. Sharpe scales with the square root of the number of independent bets, and this programme has one. §4 tested whether that could be fixed by adding markets and the answer was no: an equal-risk sleeve across ten asset classes returns a Sharpe of -0.59 against 1.50 for the single instrument. The concentration is a property of the effect, not an implementation choice.
• The signal travels; the economics do not. Across 85 independent instruments the rule beats a coin flip on mean Sharpe at p = 0.0005 — but once each instrument is charged its own measured spread only 26 of 85 remain positive and the cross-sectional mean is -0.60. Counting winners proves almost nothing here; what carries the result is the size of the effect in a handful of instruments. Anything sold on the breadth of the cross-section rather than its concentration would be mis-sold.
• One macro regime. The sample spans 2016–2026. It contains 2018, 2020 and 2022, but it is a single monetary and market era, and five-minute data does not exist to test earlier ones.
• A documented signal family. Opening-range momentum is well covered in the literature and widely traded. Crowding is a live risk. The sample shows no decay to date — the three most recent years are among the strongest — but absence of decay so far is not immunity.
• Return concentration. See §7. A meaningful share of the return sits in a small number of sessions, which converts operational downtime directly into lost edge.
• Entry is modelled at the signalling print. The backtest enters at the close of the 09:35 bar — the same price that generates the signal. A real order arrives after it, by however long the decision and the round trip take. The measured 09:35 spread is charged, so the crossing is paid for, but any drift in the seconds between the print and the fill is not modelled. This is the one execution assumption in the programme that has not been measured against the tape, and it is the natural next measurement after the stop-fill work in §5.
• Survivorship and data lineage. Bars are from a single vendor's consolidated feed, and the cross-section in §4 is built from tickers that exist today — funds that closed during the decade are absent, which biases the cross-sectional result upward by an unmeasured amount. Corporate actions on the ETFs are handled by the vendor's adjustment, not independently verified.
• The slippage measurement covers top of book only. It is valid at the sizes tested and does not extend to institutional size without the capacity work in §8.

10Status and what would change the view

Where this sits and what the next decision depends on

The programme is research-complete and deployment-pending. The parameter selection has survived combinatorial purged cross-validation with a selection bias of +0.03 Sharpe. The execution assumption that previously disqualified it has been re-measured against the consolidated tape and does not hold. The instrument selection has now survived a frozen-configuration test across 95 instruments and a 2,000-draw permutation null, p = 0.0005. What remains open is operational, not analytical — with one analytical answer that came back negative and is treated as settled: this does not scale by adding markets.

Conditions for committing capital

1. Execution infrastructure on always-on hosting with verified session coverage, given the concentration in §7.
2. Complete per-session exit recording — fill price, exit reason, stop price at fill, realised result — so live execution can be reconciled against the model continuously rather than reconstructed after the fact.
3. A rolling live-versus-model slippage comparison, with a defined tolerance that halts the programme if real fills diverge from the measured distribution.
4. Capacity work in Nasdaq-100 futures before any allocation above the ETF ceiling in §8 — now also the largest open question in §4, since the ETF proxies used there carry execution economics a futures contract does not.

What would falsify this

Live stop fills clustering materially above the measured distribution; a rolling twelve-month live Sharpe below zero; or the year-by-year surface in §3 turning negative in the most recent period. Each is monitored and each has a defined threshold.