f1b7dcf7f3
The MI sample builder used `MathAbs(labelBarOffset)` as a padding, causing rows from offset and non-offset builds to be paired with a double shift. This broke the positive control, failed the 5× gate, and voided all reported mutual‑information figures. Replace with the fixed `MiShiftPad` constant to ensure builds enumerate the same set of bars and row-k alignment is preserved. Add `BatchOptionsTotal()` to `CNeuronBatchNormOCL` and split the packed BN weight array in the learning report into separate norms for the outgoing dense matrix, gamma, beta, running statistics, and Adam moment buffers. This turns an ambiguous single‑norm reading into precise diagnostics that distinguish weight divergence from scaling issues.
69 lines
3.3 KiB
Python
69 lines
3.3 KiB
Python
"""Retail's own side looked positive once the fill was fixed. Is that the pattern or the drift?
|
|
|
|
Retrial 1 killed the fade and inverted it: on USDJPY, XAUUSD and SP500 the fade loses 0.05
|
|
to 0.14 R because retail's trend-following setups are on the right side of instruments that
|
|
rose for the whole sample. Several cells show retail's own expR positive - SP500 H1 engulf
|
|
+0.113, XAUUSD H1 engulf +0.095 - and a POSITIVE base would be better than the near-zero one
|
|
step 2 is hunting for.
|
|
|
|
Before any of that is believed it has to beat the control that makes the boring explanation
|
|
explicit:
|
|
|
|
CONTROL = random bars drawn from the SAME trend-filter state, same direction, same entry
|
|
mechanics (a stop through the bar's extreme), same stop rule, same n.
|
|
|
|
Everything is held fixed except the candlestick itself. If retail's edge is really "buy the
|
|
break of any bar's high while the 20-MA is rising, on something that went up", the control
|
|
matches it and the pattern is worth nothing. That is the same shape of control that killed
|
|
"stops are a farmable magnet".
|
|
|
|
Reported on the NON-OVERLAPPING subset, and the difference carries its own standard error
|
|
rather than being eyeballed from two columns.
|
|
"""
|
|
import numpy as np, sys
|
|
sys.stdout.reconfigure(encoding='utf-8', errors='replace')
|
|
import fills, book, retrial
|
|
|
|
SYMS = ('EURUSD', 'USDJPY', 'XAUUSD', 'SP500')
|
|
|
|
|
|
def diff_t(a, b):
|
|
"""Welch t for the difference of two independent means."""
|
|
se = np.sqrt(a.var(ddof=1) / len(a) + b.var(ddof=1) / len(b))
|
|
return (a.mean() - b.mean()) / max(se, 1e-12)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
syms = [s for s in sys.argv[1:] if s in SYMS] or list(SYMS)
|
|
print("=== RETRIAL 2: is retail's trend side a PATTERN, or just drift? ===")
|
|
print(" control = same trend state, same direction, same entry, random bar.")
|
|
print(" 'excess' is pattern minus control on independent trades - the real claim.\n")
|
|
print(f" {'sym':>7}{'tf':>5}{'setup':>8}{'dir':>4}{'k':>3}"
|
|
f"{'n':>6}{'pattern':>9}{'t':>6}{'ctrl n':>7}{'control':>9}{'t':>6}"
|
|
f"{'EXCESS':>9}{'t':>6}")
|
|
tally = []
|
|
for sym in syms:
|
|
bk = fills.Book(sym)
|
|
for tf in ('M15', 'H1'):
|
|
f = book.frame(sym, tf, bk)
|
|
for k in (1.0, 2.0):
|
|
pat = retrial.retail_arm(sym, tf, k=k, bk=bk, f=f)
|
|
ctl = retrial.retail_arm(sym, tf, k=k, control=True, bk=bk, f=f)
|
|
cmap = {(nm, d): o for nm, d, o in ctl}
|
|
for nm, d, o in pat:
|
|
c = cmap.get((nm, d))
|
|
if c is None:
|
|
continue
|
|
A = o['R'][o['indep']]; B = c['R'][c['indep']]
|
|
ex = A.mean() - B.mean()
|
|
t = diff_t(A, B)
|
|
tally.append(ex)
|
|
print(f" {sym:>7}{tf:>5}{nm:>8}{d:>+4}{k:>3.0f}"
|
|
f"{len(A):>6}{A.mean():>+9.4f}{book.tstat(A):>+6.2f}"
|
|
f"{len(B):>7}{B.mean():>+9.4f}{book.tstat(B):>+6.2f}"
|
|
f"{ex:>+9.4f}{t:>+6.2f}")
|
|
ex = np.array(tally)
|
|
print(f"\n {len(ex)} cells mean excess {ex.mean():+.4f} "
|
|
f"positive {int((ex>0).sum())}/{len(ex)}")
|
|
print(" A pattern with real content clears its control in most cells and by a margin")
|
|
print(" that does not shrink when the sample is made independent.")
|