Day-Trading Test: 53 of 58 Recipes Failed — and the Gain Came Overnight

Did trading in and out of US stocks within the same day pay in this study — a breakout out of the opening range, fading a large gap at the open, riding intraday momentum, a snap-back to the day's average price, or a sudden spike in volume? We tested five recipes on our price data.

The data base: 6,846,558 stock-days across the 4,068 US stocks in this study, drawn from five-minute price candles between 2010 and early August 2026. All five recipes, plus one add-on tested separately, were written down as exact rules before a single result was computed — pre-registered, so there was no chance to tune the rules after seeing what worked.

The result in four numbers

• 5 of 58 — only 5 of the 58 pre-registered main test cells passed our success criterion, and all 5 come from one single pattern, trading large opening gaps: four against the gap, one with it. Four of the five are short sales.
• +0.36% — the average net return of our flagship gap cell — the middle of the three rules that fade an upward gap, at 5% or more (the largest gaps paid more but occurred far less often in the period studied), where "average" means the winsorized mean: all trades summed and spread evenly, with the single most extreme 1% of results capped at each end so a handful of freak trades cannot swing the whole number.
• 39% hit rate, negative median — in that same cell, only about four in ten trades were winners, and the median trade — the typical result, half of all trades above it and half below — actually lost money (−0.73%). The positive average came entirely from a few very large winners.
• ~90% overnight — of the total price gain across the 4,068 US stocks in this study over sixteen years, roughly nine-tenths happened between the previous close and the next open, not during the trading day itself.

In one sentence: across the five day-trading recipes tested in this study, only trading large opening gaps held up in the backtest after simulated costs — four cells against the gap, one with it — and even there, the typical trade lost money, and the plus came from a few very large winners.

Where the return of the stocks in this study came from

Before testing any recipe, it's worth asking a more basic question: when did the stocks in this sample earn their return? The answer is lopsided. Across 6.28 million overnight observations, the average return (all of them spread evenly) from the previous close to the next open was +4.77 basis points, a basis point being one hundredth of one percent. The same average from the open to the close was only +0.54 basis points.

Average return per stock and day, split into the time between two trading days (overnight) and the time during the trading day. One basis point is one hundredth of a percent. Based on 6,282,496 overnight and 6,330,469 daytime observations, before costs. Source: price data, own calculations.

Put together, roughly 90% of the total price gain happened overnight, while markets were closed, and only about a tenth during regular trading hours — the window in which the recipes tested here trade. The overnight return was positive in 14 of 16 years; the daytime return in only 9. For the stocks in this sample, the gain was largely collected before the trading day began.

A separate check on 24 very large US stocks and 95,886 overnight periods traded this effect directly: buy at the close, sell at the next open. After costs it turned negative, −6.66 basis points per night, positive in 0 of 16 years — the bid-ask spread would have needed to sit under about 0.017% per side to work, tighter than the cost assumptions used here. The overnight period carried almost the entire price gain, but it could not be cut out and traded on its own.

What this is about

The question is simple: does trading within a single day, using five widely known recipes, produce a real edge once realistic costs are subtracted? Before a single number was computed, we locked in five classic day-trading recipes plus one add-on as rules, each with its entry, exit, and firing conditions fixed in advance — an opening-range breakout, a gap fade, an intraday momentum trade, a reversion trade around the day's average price, and, added later, buying right after a sudden volume spike. This series sits alongside our other five-minute studies, including the death-cross study and the crash-reversal study on 5-minute candles.

Data and pre-registered design

The core universe (the first five recipes) draws on 4,361 US stocks with a market capitalization of at least $300 million at any month-end in 2020, of which 4,068 had days that could be evaluated, survivorship-free — delisted stocks stay in the history. Each stock-day also needed an opening price of at least $5, a prior-day dollar volume of at least $5 million, and at least 60 five-minute candles (36 on a short day): 6,846,558 stock-days across 4,110 trading days. The volume-spike add-on ran its own pass — 4,361 stocks checked, 11.77 million stock-days, 788 million candles in regular trading hours (9:30 a.m. to 3:55 p.m. New York time) — narrowing to 3,219 stocks and 3,984 trading days with a candidate signal.

"Success criterion (pre-registered): a cell pays if the net winsorised mean in the main cost tier is > 0, the 90% bootstrap interval does not contain zero, AND at least 3 of 4 time slices are positive. Anything else means: the signal does not pay."

Translated verbatim from this series' pre-registered design, written down before the first result was computed.

One publisher's decision applies throughout this study and every other backtest we publish: the year 2020 is left out of every figure and every average, because the pandemic crash and rebound would otherwise distort the picture on its own. That decision is made once here and not repeated at every number below.

The main cost tier in this study was 0.05% per side, 0.10% round trip, plus a small commission simulated on every trade. Fills are conservative: a signal completing at one candle fills at the next candle's open, not the signal price. Each cell had to clear a three-part bar: the winsorized mean (the average once the top and bottom 1% of results are capped, so a few freak trades can't swing it) had to be positive; a bootstrap interval (reshuffling entire trading days a thousand times to gauge how much the result could plausibly vary) had to exclude zero; and the result had to be positive in three of four time slices (the sixteen years split into four periods, so no result rides on one lucky era). Only 5 of 58 pre-registered cells cleared all three bars; 912 cells were computed, plus 288 for the volume-spike add-on.

The one pattern that held up in the backtest: large opening gaps

A gap is the difference between where a stock opens and where it closed the day before. Our gap-fade recipe waits for the first five-minute candle to close, then decides: "go" (trade with the gap) or "fade" (bet against it), with a stop at the extreme of that first candle. Fading is typically a short sale — borrowing shares to sell first, aiming to buy them back cheaper later, betting the price falls.

Five pre-registered cells cleared our bar, and every one of them is a gap trade — four of them short sales, whose borrow fee is not included. Fading upward gaps of 3%, 5%, and 10% or more all cleared the bar in the backtest, as did fading a downward gap of 5% or more, and going with (not against) the largest downward gaps of 10% or more.

All twelve pre-registered gap rules, net per trade after 0.05% cost per side plus commission, excluding the year 2020. "Fade" means against the direction of the gap, "go with it" means in its direction. The robust average trims the most extreme one percent at each end. Bold = the pre-registered success criterion is met. Source: price data, own calculations.

The pattern in one sentence: in the backtest, after a sharp gap up the day of the stocks in this study tended to give some of that move back — fading it produced a positive robust average even though most individual trades lost; after a sharp gap down, it was the direction of that first five-minute candle — not the direction of the gap — that tended to persist. One cell needs an honest caveat: fading a downward gap of at least 5% technically cleared the bar, but the lower edge of its plausible range sat at essentially zero (+0.00005%) — a borderline pass, not a clean one. A close cousin, fading a downward gap of at least 10%, narrowly missed, with a range that still included zero.

Going with an upward gap, rather than fading it, came back negative in the backtest at all three sizes; the smaller downward-gap cells missed the bar as well. The measured advantage, where it existed, held up across every one of the four time periods we tested it in.

Robust average per trade across four periods, for the five rules that met the success criterion. All four positive means the finding does not rest on a single market phase. Source: price data, own calculations.

Broken down into the four periods, the measured gap-fade advantage showed up in every one of them, without one period carrying the whole result. That consistency is one reason the pattern survived pre-registration where the others did not; related patterns showed up in our episodic-pivots study, which also examined sharp, news-driven price jumps.

Why the typical trade still lost

This is the part that matters most about the pattern. In every one of the five cells that held up in this study, measured in the backtest, the median trade — the typical outcome, with half of all trades doing better and half doing worse — was negative, ranging from −0.63% to −1.26%. Only 36% to 40% of trades were winners at all. That means six or seven trades out of ten lost money.

And yet the average (all trades summed and spread evenly, extremes capped) came out clearly positive, between +0.211% and +0.708%. Median and average diverged this sharply because a small number of very large winning trades pulled the average far above where the typical trade landed — the profit factor (gross profit from winners divided by gross loss from losers) ran between 1.20 and 1.44, meaning winners outweighed losers in dollar terms even though they were rare.

How outlier-dependent is that? Summed by year, the flagship gap-fade cell (fading an upward gap of at least 5%) put +24.89 percentage points into 2025 alone, out of 82.3 percentage points across all sixteen years — by far its strongest single year, a clear sign that a handful of exceptional trading days, not a steady weekly edge, drove the result. The arithmetic of the backtest rested on those rare large winners: without them the rule would have been negative, and most individual trades were not among them.

The opening-range breakout

The opening-range breakout is a classic day-trading recipe, tested here in the backtest: define a price range using the first candles after the bell (the "opening range"), then buy if the price breaks above it, with a stop on the opposite side. We tested 36 variants of it — different range lengths, both long and short, with and without volume and gap filters — and all 36 came back negative.

The main cell (opening range equal to the first five-minute candle, stop on the opposite side, long positions, no extra filter) had a 35.7% hit rate, a median trade (the typical result) of −0.38%, and an average net return (all trades spread evenly, extremes capped) of −0.134%, with a plausible range that stayed entirely below zero. It was positive in 0 of 16 years, and negative before costs too, at −0.024%.

Seven of the 36 breakout variants computed, net per trade. Not one met the success criterion. The last row is not a trading rule but the plain price drift over exactly the same window, measured over the full period — and it beat every rule. Source: price data, own calculations.

The clearest sign of failure: simply buying and holding through the same time windows, no signal at all, would have averaged +0.050% — better than the breakout rule itself. Widening the range to 15 or 30 minutes, using a volatility stop, or adding a gap or volume filter changed little; every variant, including a similar recipe in our Velez open-trade study, stayed negative. Filling trades at the signal price instead of the next open made results worse still, so the failure was not an artifact of the conservative fill assumption in this backtest.

Intraday momentum and the average price

Two more recipes tested in the backtest whether a stock's early behavior predicts what happens later. The momentum rule took the direction of the first 30 or 60 minutes as its sign: a stock that rose in the morning was bought for the last 25 trading minutes of the day (3:30 p.m. to 3:55 p.m.), one that fell was sold short. Both variants averaged (spread evenly) about −0.11% net, with a 34.4% hit rate and 0 of 16 years positive — the window was only 25 minutes end to end, short enough that the roughly 0.11% round-trip cost consumed whatever the signal could contribute.

The other recipe used the VWAP, the volume-weighted average price — in plain terms, the price at which a stock traded on average that day. Betting with the VWAP trend at mid-morning averaged −0.117% net, negative in every year tested.

Intraday momentum and the volume-weighted average price, net per trade. In the reversion rules the hit rate and the median looked good while the robust average stayed negative: many small gains, a few very large losses. Source: price data, own calculations.

The reversion version — buying when the price sits 1%, 2%, or 3% below the VWAP, betting on a snap-back — mirrors the gap trades. The median (typical) trade was positive (+0.233% at 1%, +0.105% at 2%) and the hit rate was high, up to 58.4%. But the average (spread evenly across trades) still came out negative: many small wins, a few large losses, the opposite of the gap fades, where a few large wins pulled a negative median up to a positive average. One caveat: the exit carries a little hindsight (it exits at the VWAP value of whichever candle reaches it), so the result is if anything too generous — and in the backtest it still lost.

The volume spike: the clearest no

The sixth and final recipe, added after the first five, tested a simple, everyday rule: a five-minute candle that jumps at least 10% versus the candle before it, on at least triple that previous candle's volume. The rule buys at the next candle's open and sells five candles later, 25 minutes on. Only the first such signal per stock per day counted, and the signal candle itself needed at least $100,000 in dollar volume.

Starting from 9,248 candles meeting the price-jump and volume conditions, 5,016 fell out as untradable, 416 sat too close to the market close, and 255 had a data gap, leaving 3,561 tradable signals and 2,464 counted as the first signal of the day per stock, across 821 stocks and 1,458 trading days.

From raw signal to tradable trade. More than half of the candles that met the rule in the study period were not tradable — too cheap or with too little turnover. Source: price data, own calculations.

The result was the worst of any recipe in this study: a median trade (the typical result) of −3.69%, a hit rate of just 32.2%, and 0 of 16 years positive. It was negative even before costs (robust average −2.00%), so in the backtest the failure sat in the pattern itself, not in trading friction.

The volume-spike rule and its variations, net per trade after 0.05% cost per side. Not one variant was positive. The final row is the decisive control: a randomly chosen candle in the same stocks on the same days, held just as long, did markedly better in the backtest. Source: price data, own calculations.

The most damning comparison came from a random baseline: picking an entirely random candle in the same stocks, on the same days, and holding it for the same length of time averaged (spread evenly across trades) +0.62%, positive in 14 of 16 years. In the backtest, buying right after a volume spike in the stocks of this study did not just fail to make money — it did roughly 3 percentage points worse than a randomly chosen candle in the very same stock, on the very same day.

How the result develops with holding time, computed on the same set of 2,052 signals throughout. It did not improve — it got worse with every further candle. Source: price data, own calculations.

Holding longer only made things worse. Measured on the same 2,052 signals throughout, so the curve compares like with like, the average net return (spread evenly, extremes capped) fell from −1.37% at 5 minutes to −2.32% at 15 minutes, −2.74% at 25 minutes and −3.12% at 50 minutes, while the median (typical trade) slid from −1.77% to −4.70% at 60 minutes. Every sensitivity check — jump thresholds, volume multiples, all signals instead of only the first, no tradability filter — stayed negative.

What costs do to all of this

In this backtest, costs shrank the result at every level but did not decide it. Looking at the flagship gap-fade cell (fading an upward gap of at least 5%) at rising cost levels shows the simulated average shrinking steadily while staying positive at every cost level tested — before any borrow fee.

The same rule — gap up from 5%, traded against it — across four cost levels. The main assumption of 0.05% per side equals 0.10% per round trip plus commission. The profit factor divides the sum of all gains by the sum of all losses; above one means something is left over. Source: price data, own calculations.

At zero costs the average (spread evenly across trades) net return for that cell was +0.472%; at 0.05% per side, +0.362%; even doubled to 0.10%, it stayed positive at +0.261%. Every step up in cost tier made every one of the 912 core cells worse, no exceptions — a basic sanity check the design had to pass. One honest addendum: commission for the first five recipes was set at a flat 0.01% per round trip, fitting a roughly $100 stock, not the median entry price of the breakout trades actually taken, $44.35. A realistic commission would run about 0.013 points higher at the median and 0.024 higher on average — those five recipes' results are modestly too generous by that margin. The rate stayed unchanged because it was fixed before testing began, disclosed rather than corrected after the fact; the volume-spike recipe computes commission from each trade's actual entry price.

Limits of this study

The most important limitation, stated plainly: four of the five cells that held up in the backtest are short sales, meaning they profit from a falling price by borrowing and selling shares first. Stocks that just gapped up sharply can be difficult or impossible to borrow, and the fee for borrowing them is not included anywhere in this study. That fee alone could erase the entire measured advantage.

Beyond that: trades fill at the next candle's open, conservative, but real bid-ask spreads after a large gap can run wider than the 0.05% per side assumed here — the largest remaining uncertainty. The stock universe was selected by market capitalization at any month-end in 2020, which nobody trading in, say, 2012 could have known; this look-back bias is known, accepted, and affects every recipe equally, not just the winning one. Commission was understated for the first five recipes, as above. No taxes, financing costs, or market impact from a trader's own orders are included. And one stock is missing five-minute candles between June 10 and July 24, 2024, across a stock split.

The design and calculations passed all seven verification checks: exact candle counts on sample days, a short-day calendar matched against an independently built one, hand-recomputed example trades within a 0.002% tolerance, a random baseline reproducing the plain daily price drift, two independent runs producing identical result files, and cost monotonicity across all 912 core cells. The volume-spike add-on passed its own seven checks, and 21 automated unit tests passed throughout.

What remains

Across five pre-registered day-trading recipes plus one add-on, only one pattern cleared our pre-registered, cost-adjusted bar in the backtest and in the universe studied: trading large opening gaps, fading them in four of the five cells and going with them in the fifth. Even there, the typical trade lost money, and the positive average depended on catching a few unusually large winners in stocks that can be hard to borrow. The breakout, momentum, VWAP trend-following, and volume-spike recipes were all negative after costs; the breakout and the volume spike were negative even before costs, at −0.024% and −2.00%. Most of the price gain of the stocks in this study over sixteen years happened overnight, between one close and the next open — hours in which none of the day-trading recipes tested here was positioned to act. The full archive of our published backtests sits at our studies overview.

Note: This article is a retrospective statistical analysis of historical price data. It is not investment advice and not a recommendation to buy or sell. Past results say nothing about the future. Trading costs, borrow fees, taxes and the market impact of a trader's own orders can make real-world results substantially worse.