What we measured and what did not work

n = 68,918 · 2010-06-07 to 2026-10-09 · exchange data

27 markets with at least 9 complete years of history up to 2021, prices adjusted at contract rolls. Selected window = a stretch of calendar dates that closed higher in 70% or more of the last 10 complete years up to 31-Dec-2021. Each window is then looked at in every year from 2022 to 2026 (2026 through 9 October): 'positive' = it closed higher again. The comparison column is the same count for every window of the same market and the same length, without selecting. The first row counts windows; the other two count window-year combinations. Selected windows overlap heavily (a window and the one starting a day later are almost the same), so the last row keeps only those that do not overlap: its 95% confidence interval is 47.4-52.4%. Raising the threshold to 80% or 90% of years does not change the conclusion. This is what happened, not a signal.

n = 33,822 · 2010-06-07 to 2026-10-09 · exchange data

33,822 variants of mechanical session strategies on 13 markets (indexes, gold, silver, crude oil, natural gas, bonds, grains, currencies and bitcoin), each with at least 30 trades up to 31-Dec-2021 and 20 afterwards. The screen is calculated only with the trades up to 2021 and requires all of: positive net result, a sign test different from chance, and the same positive sign in two chronological splits of that period. 'Profitable in 2022-2026' = mean result per trade above zero from 1-Jan-2022 to 9-Oct-2026, after commission and one tick per side. The second row is the strategies that were net profitable up to 2021 but did not pass the rest of the screen: if passing the screen told us something, the first row should sit clearly above the second. Simulation on historical data: hypothetical results. This is what happened, not a signal.

n = 4,149 · 2010-06-07 to 2026-10-09 · exchange data

ES regular session (9:30-16:00 ET), 7-Jun-2010 to 9-Oct-2026, 30-minute period profile, prices adjusted on contract changes. At each session's close we take, for each type, the nearest unvisited level above the close and the nearest below it, among those left by the previous 250 sessions and no further than 4 ATR (20-session average true range): each session adds up to two cases per type. POC = the row with the most TPOs of a session; value-area edge = the upper or the lower limit of 70% of its TPOs; unvisited = no later session has traded that price. Single-print stretch = the longest inner run of two or more consecutive one-TPO rows; reached = the next session trades through all of it. Unfinished extreme = a high or low whose row holds two or more TPOs (no tail) and that no later session has exceeded; earlier high or low = the other highs and lows not yet exceeded. 'Reached' = the next session trades the level. The control is not a level: for each case it is the share of all the sessions before that date whose next session moved away from its own close, in the same direction, by at least that same distance measured in ATR; the column is its average. If the two columns look alike, the level does not attract price more than any other at that distance. The differences run from −1.0 to +0.7 points; over five sessions, from −0.7 to +2.0. It is what happened, not a signal.

n = 2,010 · 2010-06-07 to 2026-09-30 · exchange data

ES regular session, 7-Jun-2010 to 30-Sep-2026, 30-minute period profile. Poor extreme = the row of the session's high (or low) holds two or more TPOs, no tail. Exceeded = one of the next N sessions trades beyond that extreme. Control = the highs and lows that are not poor. Single prints = the longest stretch of two or more consecutive one-TPO rows inside the profile (tails do not count). Filled = a later session trades through the whole stretch. Control = a stretch of the same size mirrored on the other side of the close. If the level column and the control column look alike, the level does not attract price more than any other. Prices adjusted on contract changes.

n = 4,179 · 2010-06-07 to 2026-10-09 · exchange data

ES regular session (9:30-16:00 ET), 7-Jun-2010 to 9-Oct-2026, prices adjusted on contract changes. Sequence of rising lows: it starts on the session that trades above the high of the range in force without losing its low and with a higher low than the previous session's (that day already counts as the second: the minimum length is 2 sessions), goes on while no session trades below the previous session's low and ends on the session that does. Falling highs: the same in reverse. No sequence = the rest; the range in force starts with the two sessions of the day the last sequence ended and widens with each session until one leaves it on one side only. Range = the next session's high minus its low, divided by the ATR (average true range of the 20 sessions ending at the close it is measured from). Difference = the row's mean minus the mean of all other sessions. 95% confidence intervals (standard error over blocks of 10 consecutive sessions): rising lows −0.24 [−0.28 to −0.20]; falling highs +0.26 [+0.20 to +0.31]; no sequence +0.05 [+0.01 to +0.09]. With the sessions from 2022 on only: −0.22 [−0.29 to −0.15] (n = 437) and +0.23 [+0.14 to +0.31] (n = 306). It predicts size, not direction: 'closes higher' = the next session's close ends above this close, and it runs from 54.1 to 55.4% across rows (from 2022 on: 51.9% after rising lows and 54.2% after falling highs, against 52.8% for all sessions). It is what happened, not a signal.

n = 306 · 2010-06-07 to 2026-10-09 · exchange data

CFTC weekly Legacy report (futures only) and daily prices of 29 futures, through 9-Oct-2026 (from June 2010 for most; crypto from later). Extreme = the commercials' net position in the top 5% of its own recent range and that of large and small speculators in the bottom 5%, or the reverse (in equity indexes, commercials only). While the extreme is in force an entry is taken with a price confirmation rule fixed in advance, with a defined stop and exit. The result is compared with 1,000 draws of entries with the same rule, the same stop and the same holding time on dates without an extreme. 'Better than random' = positive mean result and fewer than 5% of the draws match it. 'Trades' = sum of the trades of the markets in that row. With a wider band (20%-80%) one market passes the threshold, which is expected by chance when this many markets are tested. This is what happened, not a signal.

n = 840 · 2010-06-07 to 2026-10-06 · exchange data

CFTC weekly Legacy report (futures only) for the E-mini S&P 500, prices from 7-Jun-2010 and reports to 6-Oct-2026. Reading = large speculators' (non-commercials') net position as a share of open interest, placed inside its own range of the last 156 reports (3 years): 80 or more is a long extreme, 20 or less a short extreme. 'Higher' = the close 13 weeks later is above the first close on or after the Monday following publication (the report comes out on Friday with Tuesday's data), on the continuous series adjusted on contract changes. Weeks overlap: an extreme reading lasts many weeks in a row, which is why the number of separate episodes is given; that is the real count of independent cases. 95% confidence interval for the long extreme: 51.8-69.5%. It is what happened, not a signal.

n = 52 · 2022-01-03 to 2026-10-09 · exchange data

13 mechanical rules written before looking at the period, each built from one of the platform's most prominent figures (gap fill, break of the initial range and of the opening range, return to yesterday's value, the day's objective, expected range, day of the week), applied to 4 index futures: 52 series. Regular session (9:30-16:00 ET), 3-Jan-2022 to 9-Oct-2026, on 1-minute candles; when stop and target fall in the same candle the stop is counted; costs of one tick per side plus commission. 'Distinguishable from zero on its own' = weekly t-statistic above 2 (or below -2) looking at that series alone. When this many series are tested at once, some pass that threshold by chance: the middle column says how many. The multiple-testing correction resamples 4-week blocks with the four indexes together and asks whether the best series beats what chance would give: it does not (p = 0.18). Thirteen rules do not exhaust what can be built; they say that reading these figures directly left no profit after costs. Simulation on historical data: hypothetical results. This is what happened, not a signal.

n = 14,988 · 2010-06-07 to 2026-10-09 · exchange data

Regular session (9:30-16:00 ET) of the four index futures, 7-Jun-2010 to 9-Oct-2026 (RTY from 10-Jul-2017). The day type (Normal, Normal variation, Trend, Neutral) is calculated with the report's own definition on the session cut at 12:00 and at 14:00 ET, and compared with the type the session had at the close. 'Matches' = the class in progress at that time is the same as the final one. Always naming the most frequent class would be right about half the time, so at 12:00 the class in progress says little about how the day will end. This is what happened, not a signal.

n = 4,199 · 2010-06-07 to 2026-10-09 · exchange data

ES regular session (9:30-16:00 ET). The day type compares the session's high, low and close with the range of its first hour: Normal = it stays inside that range; Normal variation = it exceeds it on one side; Trend = it exceeds it on one side by more than the size of that range and closes in the outer third of the session; Neutral = it exceeds it on both sides. 'Same direction' = the next session closes against its open in the same direction as this one. 95% confidence interval of the trend row: 46.9-53.8%. The type is only known at the close: at mid-session the class in progress matched the final one in about half of the sessions. An earlier version of this table classified the near-24-hour day, where neutral takes about three quarters; this one measures the regular session. It is what happened, not a signal.

n = 950 · 2010-06-07 to 2026-09-30 · exchange data

ES regular session (9:30-16:00 ET), 7-Jun-2010 to 30-Sep-2026. Value area = 70% of the TPOs of the previous regular session, 30-minute period profile. Acceptance = the session opens outside that area and two consecutive 30-minute periods close inside it. 'Reaches' = it trades the opposite edge of the area before the close. 'First two hours' = the acceptance completes by the fourth period at the latest. 'The opposite edge had not been touched yet' leaves out the sessions that had already traded it when the acceptance completed: in those the rule is met by definition and there is nothing left to gain. Previous-day levels adjusted to the day's contract on contract changes. It is what happened, not a signal.

n = 273 · 2021-10-08 to 2026-09-16 · exchange data

Releases from 1-Oct-2021 to 30-Sep-2026. Prior price = close of the minute before the release (8:30 ET; FOMC 14:00 ET). First-hour range = high minus low from the release minute to 59 minutes after it. Normal range = median of that same hour over the 20 previous sessions without any of the 7 releases (CPI, NFP, PPI, retail sales, GDP, PCE, FOMC); the cell is the median of range / normal range. First reaction = direction from the prior price to the close of the 5th minute; it holds if the 16:00 ET close is on the same side of the prior price (flat reactions excluded). ES front contract, CME 1-minute bars, no costs.

n = 3,769,262 · 2010-06-07 to 2026-10-09 · exchange data

24 futures, sessions from 7-Jun-2010 to 9-Oct-2026; rolling origin, years 2016 to 2026. For every report with a Monday-to-Friday breakdown two forecasts made only from the sessions before the year are compared: the rate of that weekday's cell and the rate of all days together. Both are scored on the same cases of the year with the mean squared error, multiplied by 1,000: lower is better. A positive difference = the weekday cell predicted worse than the all-days rate. Cells with at least 30 cases before and 10 in the year are included: 33,908 comparisons in the direction and continuation reports and 73,030 in the range and structure reports. The 95% interval treats each year (11 years, all markets together) as a single independent observation, because the sessions of a year share a regime and markets move together. The last column before the cases counts the markets where the cell had the larger error. 'Cases' = cell evaluations per session. It is what happened, not a signal.

n = 38,461 · 2010-06-07 to 2026-10-09 · exchange data

24 futures, sessions from 7-Jun-2010 to 9-Oct-2026. Rolling origin: each calendar year from 2016 to 2026 is taken as the test period; every report is computed on all the sessions before that year (stated) and on the sessions of the year (observed), with the product's own functions. A statement = one rate of the report, pooled over all weekdays, in one year, with at least 30 cases before and 10 in the year. Stated and observed = case-weighted means; in the families whose figures are a split (which session, which day, which side) the two match by construction and the informative column is the mean error: the absolute difference between stated and observed per statement, in percentage points. The next three columns count report-and-market combinations: 'stable' = the figure held within 5 points year after year; 'no information' = it states an edge of 2 points or more over its base rate and out of sample it did not predict better than the base; 'unstable' = the base rate predicted better, or the error exceeds what sampling explains. Of the 1,054 combinations with enough sample: 75.1% stable, 9.4% no information and 15.5% unstable; another 39 did not have enough sample. Only the reports that can be recomputed from the day-by-day series are included. Each test year loses a few warm-up sessions in the statistics that need previous days. It is what happened, not a signal.