32 tested relationships between markets, macro and capital — each with its sample, its result, its multiple-testing status and what would invalidate it. 21 of them have fallen and are published anyway: the record of what did not work is part of the product.
−0.95% per 10 bp (p=0.001)
monotonic dose-response
shift holds at 180 days (p=0.0016)
+0.16 overweight vs −0.04 for ETH (n=729 days, p=0.0003)
63.3% at 24h
67.6% vs 55.6%, median +0.83% (n=148, p=0.004)
59.4% vs 48.7%, median +0.78% (p=0.004)
54% over break-even, +138 units
35 % up at +5d/+20d vs 56–58 % baseline (in-sample, n=20)
BTC~gold +0.65 vs Nasdaq's +0.22 — gap +0.43, 96.6th percentile (p=0.011)
60.6% vs 50.1%, median +0.52% (p=0.015)
66% vs 50%, median +0.9% (n=85, p=0.001) · engine backtest 6y: +696, max DD −92, 4/4 years
fell 14 Sep: simultaneity, not lead
fell 13 Sep: does not hold outside the window it was found in
0 of 28 daily pairs show lead time
BTC vs real rate −0.28 (120 days)
+12.7 pp raw; p = 0.49 controlled
0 of 253 survive
0 of 160 tests survive
−297 actual against −1 767 inverted
59.1 % vs 53.2 % (p = 0.107)
best cell 53.7 % vs 50.7 % baseline (p = 0.38)
0 of 54 top signals survive
1 test with p < 0.05 where chance alone gives 4.2
53.2 / 45.2 / 51.5 % vs 48.4 % baseline
SOL~BTC 0.65 to 0.88 across regimes (p = 0.0003)
117 with p < 0.05 where chance gives 18 — and nothing tradable
AUC 0.498 (p = 0.94), n = 465
+2.1 pp; fails the half-split
54 % after de-duplication
fails the half-split
coin toss; half-split flips
Engine rules (thresholds, penalties, windows) are not published here — they are the mechanism, not the evidence. Methodology · Learn