Open Research
108 studies, each with its data basis, method and stated limits, including the ideas that did not survive the test. 106 rules have been through an untouched holdout so far, 2 passed. Filterable by verdict, including "no edge".
How we measure: our workflow since October 2026
One mail per published study: title, verdict, numbers, link. No promotions, no weekly digest. One-click unsubscribe in every mail.
Straight into a study
No edgeWeak Week, Friday Long After a Sweep of the Weekly Low: No Edge, No Rebound
No edgeS&P 500 Additions: Shorting From the Effective Date and the Mirrored Rule From the Announcement Both Failed in the Holdout
No edgeSingle Stocks: The Drift After Earnings Exists Before Costs, Not After Stock-CFD Costs
MeasuredRSI(2) Dip Buying on S&P 500 Stocks: Confirmed Locally in the Holdout, Not GloballyCalm uptrends eventually escalate into a melt-up, the saying goes. We measured three triggers (10-day acceleration, volatility expansion, price and VIX up together) on seven markets, search period 2000–2018. Primary rule: long after acceleration, 10 days, +38.5 bps per trade after costs (t 0.9). A random day inside the same calm uptrend earned +36.9. No trigger beats chance. Descriptive: afterwards the share of cases with a drop of 5% or more within 20 days was 16–23% against 12–15%.
Every Wednesday the US Energy Information Administration publishes crude oil inventory data. WTI jumps within seconds. Does the move of the first five minutes continue, or does it turn? 331 reports 2015–2022: going along −14.65 bps per trade after costs at 30 minutes (t −2.7), the reversal −8.87 (t −1.6). Before costs the direction is worth about −3 bps, the costs are around 12 bps. Natural gas alone costs about 35 bps per round trip. No edge.
A calm rise suddenly gets steeper and momentum builds, so jump in? On five indices (Nasdaq 100, S&P 500, Dow, DAX, FTSE) going along lost 10.05 bps per trade on average after costs in 2015–2022 (t −3.4, 141 triggers). Trading against the acceleration earned +5.24 bps (t 1.7). We ran that reversal once in the 2023–2026 holdout: −15.61 bps (t −1.6, 56 triggers). Not confirmed, no tradable edge in either direction.
An index has run at least one ATR from the open and the last stretch came accelerated: many call that the climax and bet on the reversal. On five indices in 2015–2022 the primary rule (short after the first reversal bar, 60 minutes) earned −4.36 bps per trade after costs (t −0.95, 217 trades). In 76% of the cases the climax extreme was still exceeded afterwards until the close, with the immediate fade in 94%. None of the four variants differs from zero.
“Short when the POC is lost, long when it is broken?” DAX, FTSE, Nasdaq 100 and Dow, 2015–2026, prior-day profile of the cash session. Crossing the point of control moves price to the close no differently from crossing a random level at a POC-typical position (−0.92% ATR, t −1.0 in Europe, −0.55%, t −0.6 in the US). Bounce and magnet effects are zero as well. A profile window without the New York morning, as a TradingView user suggested, does not make the levels better (DAX +0.01% ATR, t 0.0).
Does the RSI(2) pullback that passed our holdout on US index CFDs also work on individual S&P 500 stocks? Search period 2005–2018: +17.21 bps per trade after stock-CFD costs (t 2.07, 39,922 trades). Single holdout 2019–2026: +20.84 bps (t 2.00, p 0.023), confirmed locally. After the correction across all 110 holdout tests of the programme q is 0.22: not confirmed globally. The edge sits in the closing price, with entry at the next open −6.95 bps remain. In 2022 the rule lost 54 bps per trade.
After quarterly results a stock keeps running in the direction of its first reaction, the direction is real: +31.6 bps in 20 days before costs. On stock CFDs −31.2 bps remain after costs (t −3.6, 7,291 reports of S&P 500 stocks 2005–2018, hedged against SPY), because the costs are about 63 bps per trade. Short-term rules (extreme sell-off without news, gap day after results) come out between −39.3 and −2.9 bps net at costs of about 22 bps per round trip. No edge, the holdout was not run.
Index funds must buy new S&P 500 members at the close before the effective date. We measured two points in time. From the effective date, short the addition against SPY, 20 days: +90.95 bps (t 1.73) in the 2005–2018 search period, +65.27 (t 0.62) in the exceptional holdout run. From the announcement to the effective date: long lost −154.96 bps (t −5.8), the short rule mirrored afterwards looked strong at +116.87 (t 4.4) and came out at −43.71 (t −0.56) in the 2019–2026 holdout. Deletions are not reliable for lack of data.
Go long on Friday after a weak week as soon as US trading breaks the weekly low and wins it back? Nasdaq 100, S&P 500 and Dow, 2015–2022: the primary rule (week at least 1 ATR down) earned +1.67 bps per trade after costs (t 0.09, 71 trades). All eight combinations with RSI(2) and EMA filters lie between −17.6 and +5.7 bps, none above t 0.7. Friday long from the open after a weak week, without any sweep: −17.7 bps. An apparent edge of the sweep timing of +62.6 bps (t 4.2) is a selection error.
Until September 2026 we measured setups on our own calculation chain and checked them with walk-forward and Monte Carlo. A random market showed that this chain books profits even without any direction. Since then a different workflow in eight steps applies: hypothesis only with a mechanism and a counterparty, pre-registration with a SHA-256 checksum, a gate in the search period, a data check, a single holdout, broker ticks and a forward test. Of 101 rules in the holdout, two have passed.
On the three US indices, buying short pullbacks passed our holdout. Does the same family carry on the DAX, FTSE, CAC, SMI, Nikkei, Hang Seng, gold, EURUSD and bitcoin? In the 2000–2018 search period it looked that way: RSI(2) long above the 200-day average earned +22 bps per trade after costs across nine indices (t 2.3). In the single 2019–2026 holdout +6.5 bps were left (t 0.3), and none of the 23 candidates held. Plus a data error we found before the holdout: EURUSD at t 16 on free daily candles, zero on CFD minutes.
At 13:00 New York time the US Treasury publishes the results of its bond auctions. A weak auction is said to push yields up and stocks down. From 13:05 we traded the indices against a weak and with a strong auction, measured by dealer share and bid-to-cover. Result for the long maturities in 2015–2022: −2.94 bps per trade after costs (t −0.7). None of the 24 variants is meaningfully positive after costs. No edge.
When the EMA(9) crosses the session VWAP, many day traders buy or sell. On index CFDs (Nasdaq 100, S&P 500, Dow, DAX), the first cross of the day on 5-minute candles, held to the close, made +3.86 bps per day before costs in 2015–2022 (t 3.2). Costs of 2.54 bps on average leave +1.35 bps (t 1.1), positive only in the volatile years 2015, 2018 and 2020. The variant that reverses at every cross loses up to 12 bps a day. Addendum of 23 October: the test on high-volatility days only in the holdout (+8.5 bps, t 1.6) was not confirmed.
Every evening FINRA publishes which share of off-exchange volume was reported as short sales. Following the logic of the dark pool index, a high ratio stands for buying pressure. We traded the US500 CFD from 09:31 to 15:59 on the next day, long after an unusually high ratio, short after an unusually low one. Result for 2019–2022: −5.11 bps per trade after costs (t −1.0), and almost exactly the same on random markets. No edge.
At 15:50 New York time, Nasdaq publishes the imbalance of its closing auction for every stock. We sum it across 20 Nasdaq 100 heavyweights and trade the NAS100 CFD from 15:51 to 16:00 in its direction. In the 2018–2022 search period this earned +3.15 bps per trade after costs (t 2.6) and was the only test of this round to pass the gate to the holdout. In the single holdout from 2023 to June 2026 it made −0.15 bps (t −0.2). No edge.
Our first test took QQQ from NYSE Arca, where the ETF has no meaningful closing auction at all. The correction via its home exchange, Nasdaq: trading the NAS100 CFD with the published imbalance from 15:51 to 16:00 earns +1.2 to +1.7 bps per trade before costs and loses −0.76 bps after costs (t −0.7, 2018–2022). No variant is positive net, and the gate to the holdout was missed.
For SPY, the gross gain after the open rose with the size of the auction imbalance, up to +4.9 bps. The same test with the Nasdaq QQQ opening cross and the NAS100 CFD flips the sign: trading with the imbalance from 09:31 to 10:00 loses −1.9 bps before costs, and −4.8 bps on large imbalances. After costs −4.28 bps per trade (t −1.8, 2018–2022). The SPY observation does not replicate.
From 15:50 New York time, NYSE Arca publishes the imbalance of the SPY closing auction. Trading the US500 CFD in its direction from 15:51 to 16:00 lost 3.86 bps per trade after costs on large imbalances in 2018–2022 (t −3.6), and 1.51 bps even before costs. Neither larger imbalances nor exiting after the auction print help, and the same rule loses about as much on random markets. The gate to the holdout was missed.
How many SPY shares are created or redeemed each day is freely available in the issuer's fund history. The idea: after unusually large inflows the price pressure unwinds, and the same after outflows. We traded against the flow two trading days later on the US500 CFD, 09:31 to 15:59. Result for 2015–2022: −1.84 bps per trade after costs (t −0.4), three years positive, five negative. No edge.
Until 09:29 New York time, NYSE Arca publishes the imbalance of the SPY opening auction. Trading the US500 CFD with the imbalance from 09:31 to 10:00 made +0.41 bps per trade after costs on large imbalances in 2018–2022 (t 0.25). Before costs the result rises with the size of the imbalance to +4.9 bps, after costs no variant reaches t 1. Trading the auction print itself loses 4 to 5 bps because of the higher pre-open costs.
When suddenly many people read “Stock market crash” or “Recession” on Wikipedia, that is a fear signal. We evaluated the views of five such articles since 2015 and traded the US500 CFD on the day after each spike. Result for the 2015–2022 search period: +10.3 bps after costs at t 0.7 (106 days), exactly as much as the market rose on those days anyway. On random markets the same rule sits between −11.8 and +1.1 bps. No edge.
After a Bitcoin weekend the Friday level is touched again in the following week in 74.7% of cases, the equally distant mirror level in 74.3% (261 weekends, 2017–2022). Order book walls do not pull price in (38% towards, 41% away), order book imbalance earns at most +2.8 bps gross over 60 minutes against 10 bps of costs, and an estimated liquidation map hits real liquidations just as well with shifted open interest. Funding low and long/short ratio reached t 3.8 and 3.7 in 2020–2024 and −41 and −2 bps in the holdout.
Four popular trader ideas, 970 pre-registered tests on index, FX and gold CFDs over 2015–2022, and no hit after the multiple-testing correction. In expiry week the first hour after a high open runs better on average (+6.3 against +1.5 bps), not worse. The FX double top hits its target in 48.1% of cases, a random walk in 50%, also with sweep and internal sweep (49.6%), and no Fibonacci target changes the expectation (all −2.7 to −3.0 bps net). In Berlin time the 15:50 window is the best ten-minute window in the DAX gross, but at +0.7 bps it is only a third of the costs.
The expected move contained the close on 80.1% of days in 2015–2022; a random market without direction manages 80.7%. After the touch the price runs back less often than in a random market (38.5 against 42.9%), and all 15 rules against the band lose after costs. The one trace, a Nasdaq breakout long (+0.118 R), had no neighbours and earned +0.046 R (t 0.75) in the single holdout.
The assumption: after a hard run in the hour before the cash open, less comes at the open. Measured over 2015–2022 on DAX, FTSE 100, Dow, Nasdaq 100 and S&P 500, it is the other way round: after the strongest third of pre-open pushes, the range of the first hour is 19% (Europe, t 7.6) and 29% (USA, t 10.8) larger than after the weakest. It only gets smaller relative to the pre-open range (median 1.40 instead of 2.37 in Europe), and a random market without direction shows that almost identically (1.34 instead of 2.24). The hour before gives no direction, and after costs nothing is left.
On 29 September we showed that our old backtest chain booked profits even on a random market. Here is the re-run: 25 combinations of breakout rule and market, 55,337 trades from 2015 to June 2026. On the corrected chain the mean is −0.08 R per trade after costs, and no combination has an advantage over the random market. On real broker ticks none of the eleven combinations we tested reaches t ≥ 2. Six filters passed the search period, none passed the holdout.
The rule “trade against the direction on the day after a US cash-session move of at least 1 ATR” looked good in the 2015–2022 search period: +14.8 bps per day (t 2.17). Split by direction, nothing happens after up days (t 0.62, holdout t 0.15). After down days there is a rebound on average (+20.9 and +27.6 bps) that hangs on a few panic days: in the 2023 to June 2026 holdout, 9 April 2025 alone supplies 65% of the total, the five best days supply 138%, without them the mean is negative, and three of four years are negative.
A publicly shown backtest promises +982% from 2019 to 2026: if the first 5-minute candle after the open closes above the EMA(12), go long, below it, go short. We do not know its exact rules and rebuilt the obvious variants on Nasdaq, S&P 500, Dow and DAX. The order of magnitude comes from leverage times a bull market: on the Nasdaq the rule makes +129% net without leverage, “always long” in the same hours +122% (computed without costs), and with four times leverage +573% against +480%. The lead over “always long” is +2.07 bps per day (t 0.57), and on the Dow the rule is negative.
We tested nine more forced flows, from pension funds and trend followers to leveraged ETFs, a collar fund and the Shanghai Gold Exchange. The pension paper replicates for 1990–2022 (+6.95 bps per day, t 2.85); in the 2023 to 2026 holdout it makes +3.2 bps (t 0.6), and on CFDs after costs −13.9 bps. Gold's volatility jumps to 1.77 times at the Shanghai open, with no direction, and none of four older time-of-day observations held in the holdout. The FTSE event day stays the only confirmed finding (q 0.002).
We split four simple trend rules on gold daily prices from 2005 to September 2026 into four market phases. Since 2023 the 200-day average looks brilliant at +23.7% a year, but it was long 91% of the time, and buy and hold made +24.2%. In none of the three rising phases does a rule beat buy and hold, in 2013–2018 every variant loses, and in 2019–2022 the rules sit between −5.5% and +2.4% a year. The short side of the three long/short rules lost in 9 of 12 cells.
A breakout backtest on gold minute candles “earns” +0.40 R per trade on NFP and CPI days (t 2.5). A random market with the same candles earns +0.47 R. The reason is execution: the calculation fills at the level, while real market orders on these days fill on average 3.9 to 4.5 USD behind it, at a 2.00 USD spread. On real ticks +0.27 R remain (t 1.1, 49 trades), second-level rules are negative in gold even before costs (−1.4 and −1.8 bps), and Bitcoin shows the same picture.
We tested overnight breakout rules from our own rulebook: a breakout from an early 15-minute candle in Australia, Hong Kong and Japan, and in gold ahead of the Shanghai Gold Exchange open. On real broker ticks (April 2023 to September 2026) all 18 configurations lose, between −0.007 and −0.093 R per trade on average. Before the entry spread they sit between −0.022 and +0.040 R, and the spread costs 0.038 to 0.077 R. On minute data none of 9 candidates is confirmed in the 2023 to June 2026 holdout, and gold does what a random market does, minus costs.
The return of US equity indices came almost entirely overnight in 2000–2018: +3.22 bps per night (t 4.13) against −0.23 bps during the day. In the holdout from 2019 to October 2026 the day caught up (+3.07 bps by day, +4.25 bps overnight). Being long every night with a CFD costs 2 bps of spread and 1.5 to 2.1 bps of swap and ends at −0.68 bps net (search) and −0.03 bps (holdout), indistinguishable from zero. Simply holding earns more per day, in all four samples.
Larry Connors’ published rule (buy pullbacks in an uptrend) earned +48 bps per trade after costs on US index CFDs in the single 2019–2026 holdout (t 3.6) and beat random entries by +34 bps (t 2.4). In risk units it stays small: +0.07 to +0.18 R per trade, about 8 trades a year per market, and at 1% risk per trade roughly 0.6% a year to 2018 and 1.2% in 2019–2026. Stops do not help, and the three US indices are practically one signal. Caveats apply: chosen among three related rules by the highest t, a bull-market holdout, published since 2008.
We tested breakouts, pullbacks, candle patterns, calendar holding and momentum on 4-hour, daily and weekly charts with the costs of index and gold CFDs. Of 1,920 daily and weekly variants and 500 4-hour variants, none gets through the holdout and the global multiple-testing correction (best q 0.20). What earns net in the 2019–2026 holdout is mostly the upward drift: against buy and hold of the same exposure, no excess t exceeds 0.92. At holding periods of 5 to 10 days the swap eats a median 62 to 70% of the gross move, and what is significant is mostly losing.
When the VIX curve is inverted (VIX9D above VIX, or VIX above VIX3M), the VVIX is high or dealer gamma is low, the next US session is on average 0.12 to 0.35 typical daily ranges wider. This held in the search period and in the single holdout 2019–2026 (t 2.6 to 5.4), also at the same VIX level. The direction of the day does not change: long and short rules after these states sit at zero after costs. DIX, SKEW and the put/call ratio give no timing.
We put everything we thought we knew aside and searched 15 markets from scratch for intraday patterns: candles, chart formations, indicators, Al Brooks, Hougaard, the Market Wizards, smart money concepts, macro events, psychology, machine learning, brute force and forced flows. About 897,000 logged tests, 19 candidates each run once in an untouched holdout. After the global multiple-testing correction one finding remains: the FTSE on index event days. A second one, the Tokyo fix, runs as a pre-registered forward test.
When FTSE 100, STOXX and MSCI rebalance, passive funds must trade in the closing auction. A short held until after the auction print made +8.3 bps per event day in the 2015–2022 search period, +11.5 bps in the untouched 2023–2026 holdout (t 3.9, 22 of 27 days positive) and +11.9 bps on real broker ticks. It is the only pattern of our scan that passed the global multiple-testing correction. Small, rare, with clear caveats.
We ran our own breakout backtest chain, unchanged, on a random market: every minute candle mirrored around its open, same volatility, no direction. The best setup still booked +0.51 R per trade, and 393 of 498 cells showed a significantly positive result where minus costs was the only honest answer. Three sources: a volume filter that confirms late while the fill stays at the level, fills at the level on first-minute breaks, and a trailing stop simulated on 15-minute candles. After the repair the bias of the best setup falls from +0.65 to +0.05 R.
Brooks' rule of thumb that about 80% of breakouts from a range fail is right as a number: 80.3% in DAX and FTSE. A random market with the same volatility also gives 80.3%. His core setup, the second pullback at the 20-EMA, reaches its 1R target in 49.4% of trades, a random walk manages 50.1%. Of 3,586 rule variants on five indices plus gold and EURUSD, none reaches a net t of 2, the maximum is 1.95.
We combined 136 building blocks into about 682,000 simple intraday rules on eight markets and repeated the same search 20 times on random markets. The best real find has t 3.65, the best find of a typical random market t 3.81 (reality check p = 0.71). The 50 best rules confirm 0 of 50 in 2020–2022, and the one rule we flagged as an observation was negative in the holdout (t −0.59). Search long enough and you always find something with t above 3.
We tested twelve candlestick patterns (engulfing, pin bar, doji, inside bar, NR7 and more) in 31,663 variants on eleven markets, against the ordinary candle of the same colour and against a random market run with the same code. None of the pre-defined variants passes the multiple-test correction. The one striking find, a bearish engulfing in a Bitcoin uptrend with +50.4 bps and t 4.60 in the search period, returns −11.7 bps in the holdout (t −1.15).
Double top, head and shoulders, triangle, wedge, channel, trendline break and flag: 12,429 test variants on eleven markets. The classic price target comes before the stop in 46.5% of about 2.9 million bracket trades, a random walk reaches 46.8% and a random market run with the same code 46.3%. The most striking DAX flag shows t 3.56 with eight out of eight positive years, yet a random market with identical code returns +10.0 bps (t 1.86).
We tested fear spikes, AAII sentiment, put/call, geopolitics, earnings surprises, the clock change, sun and moon against 327 pre-registered trading rules. None holds: on the random market 7 rules reach a gross absolute t of 3 or more, on the real data none. The media-pessimism effect collapses from t 2.35 to 0.70 with a three-day publication lag, and the NFCI candidate (+5.29 basis points in the holdout, t 1.98) fails the global correction.
The rule of thumb “if the start is negative, the last hour is negative too” does not hold on DAX, FTSE, CAC, Dow, NQ and SPX (about 2,000 days each, 2015–2022). The first and last hour share the same sign on 48.5 to 50.8% of days, and on 49 to 51% in the 2023 to 2026 holdout. After a negative start the last hour is negative in 46 to 50% of cases, against a base rate of 46 to 51%. The rule “last hour in the direction of the first” loses 1.0 to 3.6 basis points net per trade.
Volatility-target funds should have to sell into the close after big move days. After 2-sigma days, however, the last hour in the Dow, Nasdaq and S&P 500 tends to rise: a short loses even gross (−1.2 to −4.2 bps), and bigger shocks do not help. The Japanese fiscal year-end earns +6.3 bps in a round trip around the Tokyo fix in the search period (n 48, t 2.0) and +6.7 bps in the holdout (n 21, t 1.4). Without the five best days the holdout is negative, and the same days also win in a window without the fix.
London 4pm fix, LBMA gold fixes, Bitcoin funding, the CME gap, the CNY fixing and the ECB reference rate: we tested six forced flows on minute data 2015–2022, five of them with a direction fixed in advance. In the London fix the best gross cell is +0.69 bps (t 3.5) against about 1.15 bps of cost. Bitcoin gaps of 50 bps or more fill within 24 hours 44% of the time, while the mirror level an equal distance away is reached 50% of the time. The CNY fixing shows no monotonicity with the surprise, and the ECB window is calmer than the minutes before it.
We tested 2,435 rules on intraday price dynamics: momentum, gap fill, one-minute shocks, range expansion, opening-range breaks and level touches. None passes the multiple-testing correction, and the intraday momentum of the Gao paper does not exist in 2015–2022 (−1.3 bps net). Round numbers repel price less often than placebo levels in FX and gold (43.8% against 47.5%, t −5.3). The single observation, the US reversal after a day of at least 1 ATR, made +24.2 bps in the holdout (t 1.33), 65% of it from one day.
Tight stops, adding to winners and trailing shape the equity curve, but not the expectancy. On a random market, all 3,780 rebuilt Hougaard variants end at +0.02 R gross and −0.238 R net. On real prices, moves of 0.3 to 0.5 ATR continue 52–54% of the time instead of 50%, yet 0 of 30,240 net tests reach t > 2. The one remaining lead, gap-and-go, shows +0.167 R in the search period and −0.087 R in the holdout.
Of 58,070 lead-lag cells across 15 markets, 291 reach an absolute t above 4, against 3 on the random market. The structure is real, but among these cells the largest ratio of gross return to cost is 0.58, and we require at least 2. Session handovers cannot be told apart from the random market, and the only holdout candidate earns +4.13 basis points at t 0.55.
We trained ridge regression and XGBoost on 98 lookahead-free features in a walk-forward over 79 combinations of market, time and horizon. The out-of-sample rank correlation is +0.004 (ridge) and +0.006 (XGBoost): measurably above a random market, economically zero. None of the 79 combinations reaches net t > 2, and the model rule loses 1.4 to 1.6 bps per trade on average. The rule distilled from the scan (USDJPY on the last days of the month) showed t 0.61 in the holdout.
We tested macro and central bank events (FOMC, NFP, CPI, ECB) and earnings on minute data: drift before the event, continuing or fading the reaction, announcement premium, FOMC cycle. In the 580 tests of the broad scan, 29 reach p below 0.05, and chance expects 29. The best find, gold long in the night before the US CPI, made +15.7 basis points in the search (t 3.2) and −5.7 in the holdout (t −1.1). Volatility follows the events, direction does not.
We tested 16 published rules by Raschke, Williams, Crabel, Sperandeo, Fisher and Schwartz mechanically on ten markets, 2,058 tests in the 2015–2022 search period. None passes the multiple-testing correction, 54 reach p < 0.05 against about 51 expected. The rules capture 1–4 bps gross, costs are about 2.2 bps. Crabel's NR7 breakout on gold and FX, the only bright spot (+3.49 bps, t 2.39), flips to −3.21 bps in the holdout.
A leg down, a partial retracement, and the next leg is supposed to match the first. Across 66,000 ZigZag patterns on DAX, Dow and Nasdaq at four scales, the 1.0 projection is reached before the retracement high is retaken in 44 to 47% of cases. A synthetic market built from the same bars with all path memory destroyed reaches it in 44 to 46%. No level bounces more than in the random market; where a difference is significant, price runs through the level harder. Measured moves are geometry, not structure.
The idea: if the Nasdaq trades below the prior close at 08:00 Berlin, short the first retest. We ran 1,392 variants and none is positive at t 2 or more. The primary variant loses 0.263 R per trade (t −4.95) and 0.148 R in the holdout (t −2.32). Gross, the retest is a coin flip, and the random market shows the same −0.2 to −0.3 R net.
At the overnight high, the overnight low and the prior close there are 7 to 25% more turning points than at a placebo level next to them. A random market without any directional structure produces the same cluster (19 to 22%), because price spends more time there. Per visit to the zone, price ends in a turning point in 37 to 39% of cases, at placebo levels in 39 to 40%. The holdout 2023 to 2026 confirms the picture: no level meets the clustering criterion fixed in advance.
We pre-registered seven published intraday anomalies and replicated them on index, FX and gold CFDs over 2015–2022. Gross, the effects come back at mostly 40–80% of the paper's size; net, none holds. The rule “the last half hour follows the day” earns +1.15 bps gross and −1.89 bps net (t −5.7), and ran significantly in reverse in the 2023 to June 2026 holdout (t −3.71). Pre-registered filter combinations lift the gross result to about 2 bps, none of them above costs.
We tested breakout with retest, liquidity sweeps and the ICT toolkit (order block, fair value gap, OTE, Silver Bullet, Power of Three) with 3,308 trading rules on ten markets. None passes the multiple-testing correction, and the best robust net t is 2.28. Real levels get the retest no more often than invented ones (share 0.47–0.57 against 0.47–0.56). The best sweep candidate falls from +4.18 to +0.75 bps in the holdout (t 0.34).
We tested 6,316 time-of-day and calendar variants across 15 markets. Before costs, 41 of 1,120 15-minute windows reach |t| above 3 (random markets: 1); after costs, none of the 963 windows in the ten indices is positive. The strongest FX and gold spikes (t up to 13.5) are quote mechanics at the rollover. The only stable find, EURUSD short in the London morning, made +1.86 pips in the search (t 3.4) and −0.30 pips in the holdout (t −0.5).
Japanese companies buy dollars at the 09:55 Tokyo fixing. In our data USDJPY rises into the fix on every Japanese business day (t 6.7) and falls afterwards; on holidays without a fix the picture turns around. Three rules built on this stayed positive on real broker ticks 2023–2026 (+4.5 to +5.6 bps per trade) but did not pass the global multiple-testing correction. Since 27 September 2026 they run unchanged as a pre-registered forward test until September 2027.
VWAP bands, VWAP bounces, RSI(2), Bollinger Bands, EMA crossover, MACD, Donchian-20 and ten pre-registered combinations: 3,088 VWAP tests, 5,760 indicator tests and 3,132 combination tests on eleven markets. The best net t is 2.81 (VWAP), 3.70 (indicators) and 3.35 (combinations), each time in Bitcoin, and about 4.8 would be needed after the multiple-test correction. Outside Bitcoin no test reaches t 3. Four indicator tests and one combination test reach t 2, the random markets deliver five and two.
Whoever has traded a move from the start sees the pullback differently from someone who missed it: later re-entries or none, but adds to the runner. The market does not know who was in. This is the pre-registration of a two-part study: what a late pullback entry is objectively worth on eleven years of index data, and how the trader's own state shifts the entry threshold, measured on twelve real chart situations in four states, from flat to +5 R open profit. Data collection is open; the result is published whatever it shows.
A language model writes an expert advisor in minutes that compiles and wins in the backtest. The mistakes are rarely in the syntax. They are in what the code assumes about the market: information that did not exist at entry time, fills that never happened, costs that are missing, session hours that are right until the clocks change. Twelve error classes, each with the number from our own tests or from a client case, and the checklist against them.
Gao, Han, Li and Zhou report that the first 30 minutes of the session predict the last 30 (SPY 1993–2013, R² 1.6%). On index CFDs 2015–2026 the slope is zero or negative on Nasdaq, S&P and Dow, and significant on DAX (t = 4.2) and FTSE (t = 6.1). Remove the overnight gap from the predictor and every slope goes to zero. What survives is gap persistence into the close, worth 0.6 to 1.3% of variance, shrinking out of sample, and the timing strategy loses 1 to 3 basis points a day net in all five markets.
Zarattini and Aziz report +0.13 R per trade and a 24% hit rate for a 5-minute opening-range breakout on QQQ, without slippage. On Nasdaq CFD data 2015–2026 we get +0.131 R and 23.2%, the paper reproduced to the second decimal. Net of spread and slippage the same rule makes +0.002 R on NQ and −0.04 to −0.08 R on the other four indices. The first candle does carry about +0.1 R of direction information against a random-direction benchmark; costs take all of it.
Every trade with a stop and a target has a hit rate a coin flip would also produce: stop divided by (stop plus target). 67% on a 2:1 bracket is exactly chance. 35,000 coin-flip trades per market on DAX, NQ and SPX confirm the formula within one or two percentage points, the expectation stays at zero, and whoever counts only 'completed trades' measures their holding time, not their strategy.
Doji, opposite-colour candle, break of its extreme — on a 2-minute chart it looks like trapped traders. Across 27,425 patterns on DAX and FTSE (2015–2026) it returns −0.186 R net, exactly as bad as the benchmark 'every opposite-colour candle' (−0.184 R). The doji carries zero information, and costs eat 31 to 43% of the risk.
Six breakout setups on DAX, FTSE, Dow and Nasdaq, unselected, 66,511 trades from 2015 to June 2026: +5,243 R. The number is correct and misleading. Three of those years were dead (+15 R, drawdown −449 R); the remaining eight deliver +621 R per year at −136 R drawdown. What counts as a bad month depends on which of the two distributions you ask.
An exit that flattens the position on the second counter-candle on M5 sounds like the clean answer to profits given back. Across 66,511 trades and eleven years it loses to the trailing stop in every market and every setup: paired −0.072 R per trade (t = −15.8). An observation month that seemed to confirm it hung on a single trend day.
Intuition says big trend days come from news. Across 2,949 trading days and 66,511 trades we do not find that. Only 34% of the 50 best book days fall on event days or the day before one; the base rate is 29%. The calendar works for excluding — FOMC days push the US session to −0.10 R/day against +1.19 on quiet days (t = −2.9) — not for collecting.
Across 171,180 index trades, breakout setups run at +0.149 R in the top volatility tercile and +0.046 R in the bottom one. Fade setups flip the picture: +0.130 R in low, +0.099 R in high volatility. The regime is known before the open, but honestly counted it costs more than it earns: as a filter it cuts max drawdown from −449 to −259 R for 17–19% of return — insurance, not edge.
Hammer, doji, fat body, long wick: we put the shape of the reference candle against the success of the breakout, across 4 million breaks on gold/FX and 103,182 momentum trades on our episode basis. Result: the close relative to the break side separates −0.115 R from +0.135 R (t = +24.2), body ratio explains practically nothing within the close classes, and the weekday does not survive out-of-sample.
The first 5-, 10- or 15-minute candle after the cash open matches the day's direction on 58 to 64% of days — with t-values of 9 to 15. It sounds like a forecast, but it is arithmetic: the candle is part of the day. Measure the day from the candle's end and 49.6 to 52.6% remain — a coin flip in 11 of 12 market-window cells. Only the 15-minute candle on NQ shows a weak residual.
We ran our setup family on three additional indices. The SMI loses −0.098 R net (n=20,433), the CAC 40 keeps +0.078 R and still makes the portfolio worse, and the Canadian CA60 has a raw edge of +0.172 R that flips to −0.138 R after costs. The metric that decides is not the spread in points but the spread as a share of the risk.
The low gets taken out once more after entry — so the level is known and gets deliberately swept? Across eleven years and five indices the data show the opposite: the stop rate is falling, also relative to every other setup in the same market. A 5-minute close separates +0.409 R from +0.006 R (t = +27.5, OOS-stable) — and still cannot be traded.
A strong push between 08:00 and 09:00 Berlin says nothing about the day's direction: 51.5% of days close in the push direction, 69% erase it completely. As a filter on 41,668 setup episodes it still works — trades with the push +0.199 R (t = +6.6), against the push +0.075 R (t = +0.3), baseline +0.122. The difference holds in both halves of the sample and in 17 of 18 setup cells.
Two 10-minute reference candles around US midday (18:10 and 18:30 Berlin), breakout in both directions, stop at the opposite side. Across roughly 8,800 first-break trades per window on NQ, Dow and S&P (2015–2026) both are negative: −0.106 R (t = −6.9) and −0.091 R (t = −5.5). And the direction of the first break says nothing about the second: 52.6% versus 52.0%, chi-squared = 0.2. The first attempt showed +0.90 R — a volume-gate artefact.
A breakout setup in the US midday lull showed +0.90 R at a 60% win rate — it would have been the best setup in our system. Without the volume confirmation filter: −0.11 R at 29%. The filter had shifted the trigger by a median of 100 minutes while the entry stayed booked at the old level. A case study in a lookahead trap that any backtest with volume confirmation can reproduce.
We ran the reference-candle breakout across 30 time cells on EURUSD, GBPUSD, USDJPY and gold, net of commission. The London morning window loses everywhere (EURUSD −0.152 R, t = −8.3; n ≈ 2,900 per cell), and so does its mirror image as an anti-breakout (−0.186 R). Only one window holds, in the late London morning: USDJPY +0.39 R (t = 15), gold +0.26 R (t = 11.5), in both sub-periods.
When the first 5-minute candle after a breakout entry is a pin bar with its wick pointing in the break direction, the trade ends at −0.599 R on average instead of +0.251 R (n=11,714, t = −31.9). Yet no reaction to it helps: closing immediately costs 0.023 R per trade at portfolio level, an earlier break-even adds +0.001 R. The pin bar is a proxy for where the trade stands after five minutes, not a shape effect.
The Dow is price-weighted: its two most expensive stocks make up 20.7% of the index, the world's most valuable company only 2.4%. That makes it a financials-plus-industrials index — and it shows in the calendar. FOMC moves Dow and NQ almost equally (range +0.42 and +0.36 ATR), but the setup expectancy only flips on the NQ (−0.219 R, t = −3.9). More movement is not more edge.
The US session is strong, the FTSE had a good push into the second half of the day — is it worth holding the trade past the London close until the US cash close? At the day level this yields +1.10 points (n=500, t = +0.76); at the trade level the question is almost moot: with a trailing exit, 49 of 22,002 episodes are still alive at 16:30 London. And the one cell that looked naively significant (t = +2.54) drops to t = +1.57 after day clustering.
The entry point for everyone who wants to trade automatically or already does: what a trading algorithm really is, which families exist and what kills each of them, what is realistically achievable — and the four traps where most systems fail before they ever go live.
Europe closes at 11:30 New York while the US session runs four and a half hours longer. When the NQ is down 1–2% by then, a full V-reversal follows on just 7.8% of days — and the harder the sell-off, the rarer the reversal, not the more likely.
123,193 trades across 10 markets and 11.4 years, split into structural edge, market regime and fresh anomaly. The structural edge carries over five years (β=0.90). A fresh anomaly loses 89% of its size in a single half-year. The past is a forecast — just not the part everyone watches.
Four context features in our episode base contained information that did not yet exist at entry time. 60,129 of 123,193 episodes were affected. A prediction model's tercile spread fell from +0.726 R to +0.191 R after the fix. How we found, proved and closed the leak — and why the models trained on it cannot be salvaged.
Indices rise over the long run, so intraday long setups should have a tailwind. Across 123,193 trades they do not: long +0.077 R vs short +0.081 R. The reason is a timing mismatch — 62 to 85% of the upward drift happens overnight, in the hours when intraday setups are not in the market.
Two indices of the same pair — DAX and FTSE, Nasdaq and Dow — show the same pre-open reference candle colour on 70% of days. The obvious hope: colour agreement as confirmation, a colour split as a warning. Across 7,108 continuation episodes the aligned-minus-split difference sits at |t| ≤ 1.0 in all four cells, with point estimates leaning towards split. The partner candle is redundant information — the third failure of cross-market conditioning.
Gamma wall, previous-day high/low, overnight range, opening-range line: we coded mechanical take-profits at visible levels and paired them against the trailing stop on identical trades. Across four tests the result is uniform: never better, mostly clearly worse (t = −3.1 to −21.9). The reason is the right tail — the 6.6% of trades reaching ≥ 3 R carry the system, and a level TP cuts exactly those.
Two European indices, 73% correlation over the first 90 minutes. When one pulls away, the other should follow. Across 2,892 shared trading days it does not: the DAX's rest-of-day return after the largest FTSE lead is −0.030%, after the largest FTSE lag it is +0.063% — the gradient points the wrong way, and no quintile reaches |t| ≥ 2. On a daily basis across five indices there is no catch-up either.
A bias derived from the open relative to the pivot range (Fisher/ACD logic) predicts very well which prior-day extreme the market will reach: at ±2 it is 69% against a 37% base rate. As a trade filter the same bias adds nothing — with vs against Δ = −0.001 R at n = 38,260. Once the extreme is reached the information is spent: 50/50 in every bias state.
12,949 mechanically defined 1h pin bars on DAX, FTSE, NQ and DOW, traded against the wick: −0.049 R, the same as any arbitrary candle with the same bracket geometry. At the 12h extreme too: −0.057 R. Only the second 1h close beyond the pin extreme lifts the result to +0.009 R (n=4,558, t ≈ 4 against the baseline) — real, but only cost-covering. The reason: activation structurally turns the reversal into a break.
After an activated double bottom, does price reach the 1.618 extension before breaking the low? In 67% of 26,795 sequences on DAX, FTSE, NQ and DOW — a driftless random walk would have done so in 71% (t = −13.5). The 0.5 pullback entry wins 35.5% against a 34.8% random-walk expectation, +0.020 R. And the distribution of where runs end shows no hump in the 1.618–2.0 zone. The zones forecast nothing the geometry does not already dictate.
Intuition says: once an early setup has exposed the day as choppy, the later ones should be skipped. Measured, the opposite is true. The late breakout setups of the session run markedly better after an early whipsaw — US +0.024 → +0.335 R (t = +5.1), EU +0.111 → +0.533 R (t = +6.1). A day lockout would have removed exactly the best trades.
When a reference range breaks up first and then down, the second break is negative in all four breakout setups and both market groups: US between −0.071 and −0.377 R, EU between −0.028 and −0.315 R. Only the first break carries the edge. Trading both sides flips the US book from +1,949 R to −839 R.
Gamma flip, call wall, put wall: nearly every freely available GEX map blends all expirations into one number — and ends up dominated by 0DTE gamma that physically decays during the session. A live decomposition on NQ (2026-06-10): how a call wall 'migrates' 1,300 points intraday without a single position changing, why intraday recomputation is repainting — and how a three-layer model (structure 7–45 DTE, near OI 0–5 DTE, 0DTE volume) nailed the day's high to the point.
The first edge in the series that doesn't come from a price pattern but from a reason: forced participants (inflows, fund rebalancing) around the month boundary. Tested cross-asset (DAX/FTSE/NQ/Dow, 2018–2026, net of costs). The accumulated return concentration collapses under a clean significance test — but the trader frame reveals the actually usable part: EOM days are 4/4-significantly larger movement days (extra movement at open & close), more breakout-friendly on 3/4 markets. A case study on how the measurement frame decides between a find and a null.
The process document for the whole series: the complete workflow from idea to MT5 algo in five phases, illustrated by the cross-asset 2nd-candle run. With the guiding principles that separate real from apparent edges — and why only one market survives in the end.
Trailing, break-even, partials, multi-TP, exit at the Asia/ONR edge — what actually helps? Cross-asset study (DAX, FTSE, Nasdaq, Dow) over ~2,090 days per index, 37 exit variants tested net-of-spread against a reference rule.
The validated system as a deployable MQL5 algo — plus the real MT5 run against buy-and-hold and the perfect exit. NQ beat buy-and-hold ~2× (through end-2024), DAX is dead. And the most honest chapter: edge perishability — the backtest stayed green while the real account lost from 2025.
Phase 4 of algo development — only on the validated core (NQ, DAX). Optimal-f (Kelly), fractional Kelly, vol-targeting and max-DD-stop, computed on out-of-sample parameters. Why full Kelly is suicide and quarter-Kelly the floor.
The stress-test paper: out-of-sample split, rolling and anchored walk-forward, Monte-Carlo (shuffle + bootstrap) and deflated Sharpe on the final ruleset. Cross-asset over ~2,090 days per index, net of spread. Only one market passes.
How far does the DAX break out of the 08:50–09:00 reference candle? A statistical study across 2,070 trading days (2018–2026): range, reversal, timing, offset and Asia/overnight sweeps — no strategy recommendation, just the data.
Does the DAX reference-candle behaviour transfer to other indices? The same methodology applied to FTSE 100, Dow Jones and Nasdaq 100 over ~2,090 trading days per index — the answer is consistent.
Does the 2nd 15-minute candle after cash open deliver a tradeable trend breakout? A cross-asset study (DAX, FTSE, Nasdaq, Dow) over ~2,090 days per index — tracking the move to the close instead of to the first opposing candle.
Does the 7th 15-minute candle (~90 min after open) deliver a tradeable trend signal? A cross-asset study (DAX, FTSE, Nasdaq, Dow) over ~2,090 days per index — directly comparable to the 2nd candle.
A technical analysis of what machine learning actually delivers in trading – and where the promises don't hold up to reality.
How do you distinguish robust strategies from lucky flukes? A deep dive into the methods used by professional quants – from historical origins to modern application.
Mechanism edges after the fresh scan: FTSE index event days are confirmed, the Tokyo fix runs as a pre-registered forward test until September 2027. Open: data after 5 June 2026 as a fresh, untouched sample for both, plus the futures roll, which the scan did not cover.
RSI(2) pullbacks on US index CFDs: the rule passed the holdout and the global correction, and in MetaTrader on broker ticks it delivers the same as the research. What is open is the forward test on data that did not exist at test time. On nine indices outside the US the same idea did not hold in the holdout.
CAC 40 and SMI on index event days: the same rule as in the FTSE, same direction in the holdout but not confirmed (q 0.18). Data after 5 June 2026 will be the next untouched sample.
Missing a topic? Submit your own question →
Weekly notes on TradingView
Not studies: the reading of the week with confirmation, invalidation and limits, in the same format as here.
All notes on TradingView →An observation from your own trading, a rule from a video, a rule of thumb you do not trust: hand it in here, with a chart, a rule set or data attached. It is measured with the same methodology as everything on this page, and a negative result is published just like a positive one.
Honest expectation: there is no commitment to take it on and no date. What gets measured depends on whether the question is stated so it can be falsified, and whether the data for it exists. Anyone who needs a binding verdict by a date takes the strategy validation as a paid project.