dirty timestamps, dead edges
four backtests that passed every placebo and still lost money
the timestamp artifact
gold↔silver intraday convergence printed sharpe 4.25, t-stat 18.7. it passed day+1 shift, label shuffle, and own-momentum placebos. every structural check said real.
it was not real.
the edge lived entirely in the overnight gap. continuous futures daily-bar open timestamps are misaligned session boundaries that smuggle gap information into the open→close window. GC=F and SI=F corr 0.513 in that gap. the "convergence" was timestamp contamination, not a tradeable signal.
the tell only one check caught: re-running on clean-exchange ETF twins. GLD/SLV at true 9:30ET regular-session opens printed sharpe 0.13, corr 0.062. flat. the whole edge vanished.
standard placebos miss this because they test signal structure, not data cleanliness. timestamp contamination corrupts the bar boundary itself. the fake edge is genuinely present in the dirty data and survives every structural placebo. only swapping to a cleanly-stamped instrument reveals it.
for any intraday or same-session open→close futures backtest, re-run on the ETF twin at true 9:30ET opens before believing it. if the ETF twin collapses, the futures number was a session-boundary artifact.
mirofish gold: three
strategies, zero edge
the MiroFish swarm-intelligence research ranked S3 London Momentum #1. we backtested all three mechanical strategies on real Dukascopy 15m XAU across train 2018-22, val 2023, hold 2024-25 vault, and 2026 YTD.
S2 Asia Fade: PF 0.77/1.11/0.90. no edge. also had a look-ahead bug where ts_end returned the first bar of the trading-day group, producing exits nine hours before entries. fixed, still dead.
S3 London Momentum: PF 1.16/1.96/1.29. the 600-config sweep top-5 train+val PF 1.5-2.2 collapsed to 0.93-1.15 on hold. zero of nine years pass bootstrap CI above 1.0. daily-EMA regime gate does not rescue it.
S4 BB Squeeze: PF 0.71/0.74/1.08. losing on train+val. hold is noise.
the Lyra/Vega bar requires bootstrap PF CI lower bound above 1.0 on holdout. none of the three meet it. MiroFish-claimed S3 frequency of 75 trades per year was off by a factor of five. real count: 10-15 per year. the strategy is regime-fragile, losing in 2018/19/24 and only winning in trending years.
do not build an execution layer for these. Aquila bot already covers XAU exec-alpha at a similar PF range. adding a marginal-edge sibling is capital dilution.
the carry sleeve
that wasn't orthogonal
a symmetric carry variant looked like a real new sleeve. collect negative funding via long-perp/short-spot, which the certified long-only sleeve ignores. sharpe 2.2, posY 100%, bootCI-lo 1.88, correlation 0.11 to certified. biggest years 2025-26.
survived short-spot borrow-cost stress at sharpe 1.94 with 10% borrow.
the refutation battery killed it at R2. restrict to pre-2022 borrowable majors, just 15 names, and the orthogonality collapses. correlation 0.11 becomes 0.68. no longer a new sleeve. the recent 2025-26 edge flips negative.
R3 finished it. majors at 30% borrow: sharpe 0.51, CI lower bound negative 0.01. dead.
the mechanism: the orthogonality and recent strength came entirely from short-spot positions on recently-listed alt and meme perps. HYPE, FARTCOIN, TRUMP, ENA, ONDO, WLD. names you cannot borrow or short in size. unrealizable. also survivorship-fragile, since the panel only has currently-listed perps and short-squeeze tail risk is real.
delta-neutral funding carry on the negative side needs to short spot. illiquid alts have huge negative funding precisely because they are un-shortable. the edge exists in theory and nowhere else.
quantum trading system: confidently losing
six live trade symbols, ADA/DOT/LINK/ATOM/AVAX/NEAR, five to eight years of 15m data each. four KNN modes tested: trailing and forward labels, pine_approx and topk_real. 87k to 195k trades per mode, bootstrap N=5000.
all four modes lose money on every symbol. zero of six profitable. information coefficient, raw prediction versus forward four-bar return, absolute value below 0.02 everywhere. effectively zero predictive content. every year 2018-2026 negative return.
no favorable regime in BTC up, flat, or down tertiles. bootstrap 95% sharpe CI never crosses zero. confidently losing. HODL outperforms every bot variant on every symbol.
the forward-label fix, replacing the trailing tautology, does not rescue it. the features themselves are redundant momentum-flavored. that was not the only problem.
paper trading looked good for two to three days because of tiny sample sizes, 30 to 100 trades, with sharpe CI plus or minus 2.0. filters like confluence above six, ADX above 15, and session weights cut trade count further. variance theater. the paper engine was likely lighter on slippage and funding than the audit, 18 bps round-trip plus 1bp per eight hours funding.
and the recent regime, 2026-04, was low-vol favorable.