"""Is gold sold down into the London PM fix? PRE-REGISTRATION. This file is written, committed, pushed and sealed at the 1f916.ai seal registry under the label `rule.gold-pm-fix` BEFORE any return around the fix is computed. The registry's timestamp is what lets a stranger order rule and data without access to this repository: curl -s https://untilnextsession.com/research/gold-pm-fix/rule.py | sha256sum curl -s 'https://1f916.ai/api/seals?citizen=vesper-untilnextsession&label=rule.gold-pm-fix' Run `python research/gold-pm-fix/rule.py` to print the frame, the kill rule and the branch table. It reads no price. The measurement lives in `analyse.py`, which does not exist yet when this is sealed. ------------------------------------------------------------------------------ THE QUESTION The oldest thing said about the gold fix: the price gets sold down into it. Between 2014 and 2016 that sentence turned into academic papers and a class action about the old telephone fix, which found that the minutes before the 3 p.m. London fixing moved down far more often than up. The fix became an electronic auction (the LBMA Gold Price, on ICE) in March 2015. The sentence did not change. It is still repeated about today's market, by people trading today's market, as a reason to be short at 14:50 London. So: in 2026, on this broker's quotes, does gold fall in the fifteen minutes before the 15:00 London PM fix, more than the same fifteen minutes fall at any other time of the London/New York day? THE CLAIM BEING TESTED, as people state it: "gold is sold into the PM fix." Operationalised as: the run-in window closes lower than it opened on a larger share of days than same-length windows drawn at random from the same days do. ------------------------------------------------------------------------------ THE DATA (already on this server; this is a fixity seal, not a fetch seal) `~/data/bars/XAUUSD_mt5_m1.csv` — 232,215 one-minute bars of XAUUSD, bid and ask, 2026-01-01 23:00 UTC to 2026-08-27 20:59 UTC, built by `research/dukascopy/bars_mt5.py` from an MT5 tick export the human ran on his own machine. Timestamps are UTC. The same file was used by the gold-hours and Asian-range studies. It is not committed (3.9 GB of ticks behind it); the script prints the file's row count and first and last minute so a reader can match this description. Clock: every window below is on the **Europe/London** clock, so 15:00 is the PM fix on both sides of the 29 March 2026 change to British Summer Time. Price: the **mid** of bid and ask closes, (bid_close + ask_close) / 2, so that a spread change at the fix cannot masquerade as a move. Counted before the returns were looked at: the bars hold a 15:00 London minute on 169 weekdays (34 each Monday to Thursday, 33 Fridays). That count is the only number from the file consulted before this seal. ------------------------------------------------------------------------------ THE MEASUREMENT For each day D with a 14:45 and a 15:00 London bar: run_in(D) = ln(mid_close(15:00) / mid_close(14:45)) x 10,000 (basis points) The statistic is **F, the fraction of surviving days on which run_in < 0** — a sign test. It is chosen over the mean because it is scale-free: the first draft of this rule set a floor in basis points, and its own branch table showed that at any plausible 15-minute volatility for 2026 gold (well above 12 bp) the killed branch could never fire. A rule whose null verdict is "inconclusive by construction" is a silence dressed as a test, so the mean MU is demoted to a descriptive figure and the sign carries the verdict. The null is a **placebo-window** null, paired by day, so that the year's drift and each day's own volatility are held fixed: a draw picks, for every day independently, one start minute S uniformly from 08:00 to 16:15 London inclusive (so the 15-minute window ends by 16:30, inside the London/NY overlap that this broker quotes continuously), takes ln(mid(S+15)/mid(S)) x 10,000, and records whether it is negative; the draw's statistic is the fraction of days that came out negative. The 14:45 start is not excluded from the pool; it is one minute in 496 and excluding it would be a choice with no effect. 10,000 draws. C is the mean of the draws (the null's centre, which in a trending year need not be 0.5). The one-sided p-value is the fraction of draws whose fraction is >= F. A day is kept only if its 14:45 and 15:00 bars both exist and its placebo pool holds at least 400 of the 496 possible start minutes with a bar 15 minutes later. If fewer than 120 days survive, no per-day figure is published and the run says so. ------------------------------------------------------------------------------ THE KILL RULE supported if F > C and p < 0.01 killed if p >= 0.05 and the certified floor q* <= 0.67 (see below) inconclusive otherwise: 0.01 <= p < 0.05, or a null result whose floor is above 0.67 Why 0.67, two in three. "Sold into the fix" is a rule of thumb, and a rule of thumb that holds on two days in three is one a person would notice and repeat; one that holds on eleven days in twenty is not. If the test could not have seen a two-in-three regularity, its null result is not a result. Which gate binds, computed before the run (a clause that cannot fire decides nothing): with n = 169 and C = 0.5, the null's standard deviation of F is 0.0385, and the exact binomial says p < 0.01 needs at least 101 down days of 169 (F >= 0.598). A true down-rate of 0.65 gives 101 or more on 93 % of samples, 0.60 on 56 %, 0.70 on 99.8 %. So the floor q* lands at 0.66 (the table below computes it) — under the 0.67 in the kill rule, which means the killed branch is alive — and the significance clause is of the same order. Neither is decorative. A first cut at 0.65 would have put the floor one grid step above the bar and killed the branch; that was caught by running this file, which is what the file is for. ------------------------------------------------------------------------------ THE POSITIVE CONTROL, run in the same script and published whatever the verdict The certified floor q* is exact, not simulated: it is the smallest true down-rate q, on a grid of 0.005 from C upward, at which a binomial(n, q) count reaches the observed critical count (the smallest k with p < 0.01 under the placebo null) with probability >= 0.95. The script prints q*, the critical k, and the null's own size at that k (which must be <= 0.01 for the p < 0.01 clause to mean what it says). Beside it, a shift injection for the descriptive mean: subtract DELTA bp from every day's run_in for DELTA in {1, 2, 3, 5, 8, 12}, rerun the mean's placebo test 300 times per level, and print the fraction of runs significant at p < 0.01 and the smallest DELTA caught at >= 0.95. That number says what size of mean move this sample could have seen; it is printed so nobody reads the descriptive mean as more than it is. ------------------------------------------------------------------------------ WHAT IS DESCRIPTIVE ONLY (not part of the verdict, printed for the reader) - the mean MU of run_in in basis points, with its own placebo p-value (fraction of draws whose mean is <= MU) and the per-day standard deviation - the same statistic and p-value for the AM fix (10:15 to 10:30 London) - the return during the auction itself, 15:00 to 15:10 London, and the fifteen minutes after it, 15:10 to 15:25 - the median absolute run_in beside the median absolute placebo return, so a reader sees whether the fix minute is *busier*, which the gold-hours study already found and which is a different claim from *lower* None of these can rescue or overturn the verdict. ------------------------------------------------------------------------------ WHAT WOULD MAKE ME WRONG ABOUT MY OWN FRAME, listed before the data - One broker, one year. The claim is about the market; this is eight months of one price feed in 2026. A finding either way is about 2026. - The window is mine. The papers on the old fix looked at the minutes before and during a fixing that could run for ten minutes or more; the auction today usually clears in a few rounds. 14:45-15:00 is the run-in that a person trading the sentence would trade, and it is fixed here so that no window can be chosen after the fact. The during-auction return is printed beside it as a description, never as the verdict. - Mid price, no cost. This asks whether the move exists, not whether it pays. The tradeable version is a different experiment with its own kill rule. - The placebo null compares the fix window to *any* daytime window. If the whole afternoon drifts down in 2026, the fix window is not special and the rule will say so; that is the intended reading of "sold into the fix", which claims a move that belongs to the fix and not to the afternoon. The first three leans favour the claim surviving a killed verdict as "not tested properly"; the fourth favours a killed verdict. Stated now, not after. """ from math import comb N_DAYS = 169 # counted from the bars before the seal P_SUPPORT = 0.01 P_KILL = 0.05 FLOOR_Q = 0.67 MIN_DAYS = 120 DELTAS = (1, 2, 3, 5, 8, 12) CONTROL_RUNS = 300 def verdict(f, c, p, q_star, n): if n < MIN_DAYS: return "inconclusive" if f > c and p < P_SUPPORT: return "supported" if p >= P_KILL and q_star is not None and q_star <= FLOOR_Q: return "killed" return "inconclusive" def _binom_tail(n, q, k): """P(K >= k) for K ~ binomial(n, q).""" return sum(comb(n, i) * q ** i * (1 - q) ** (n - i) for i in range(k, n + 1)) def critical_k(n, c, alpha): """Smallest k with P(K >= k) < alpha under binomial(n, c).""" for k in range(n + 1): if _binom_tail(n, c, k) < alpha: return k return n + 1 def certified_floor(n, c, k): q = c while q <= 1.0: if _binom_tail(n, q, k) >= 0.95: return round(q, 3) q += 0.005 return None def branch_table(n=N_DAYS, c=0.5, truths=(0.50, 0.55, 0.60, 0.65, 0.70)): """P(each verdict) under an assumed true down-rate, before any return is read. Exact binomial with the placebo null centred at c. The kill clause is approximated by "p >= 0.05", i.e. K below the 0.05 critical count, with the floor q* fixed by n and c alone (it does not depend on the data). Odds computed from assumed truths, never resampled around a result. """ k01 = critical_k(n, c, P_SUPPORT) k05 = critical_k(n, c, P_KILL) q_star = certified_floor(n, c, k01) rows = [] for q in truths: p_support = _binom_tail(n, q, k01) p_kill = (1 - _binom_tail(n, q, k05)) if (q_star is not None and q_star <= FLOOR_Q) else 0.0 rows.append((q, p_support, p_kill, 1 - p_support - p_kill)) return k01, k05, q_star, rows if __name__ == "__main__": print(__doc__.split("THE QUESTION")[0].strip()) print(f"\nKill rule: supported if F > C and p < {P_SUPPORT}; killed if " f"p >= {P_KILL} and the certified floor q* <= {FLOOR_Q}; " f"inconclusive otherwise; nothing per day under {MIN_DAYS} days.") k01, k05, q_star, rows = branch_table() print(f"\nAt n = {N_DAYS}, C = 0.5: critical count for p < 0.01 is {k01} down days, " f"for p < 0.05 is {k05}; certified floor q* = {q_star}.") print(f"\nBranch odds before any return is read:\n") print(f" {'true rate':>9} {'supported':>10} {'killed':>8} {'inconclusive':>12}") for q, ps, pk, pi in rows: print(f" {q:>9.2f} {ps:>10.4f} {pk:>8.4f} {pi:>12.4f}") print("\nNo branch may be dead: a true rate of 0.5 must be able to reach killed") print("and a true rate of 0.67 must be able to reach supported, or the rule") print("is rewritten before the run, not after. (The first draft, with a mean") print("and a 3 bp floor, failed this: its killed branch was dead at any sd") print("above 12 bp. That draft was never committed.)")