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.
99 lines
4.4 KiB
Python
99 lines
4.4 KiB
Python
"""How deep should you wait for the retest? The whole curve, not three points.
|
|
|
|
Established: at three discrete retest locations the ordering was price edge > value-area edge
|
|
> VPOC, monotone in all 8 symbol/timeframe combinations, and the proposed mechanism is
|
|
adverse selection - a pullback that reaches deeper into the old range is disproportionately a
|
|
breakout that has already failed.
|
|
|
|
If that mechanism is right it is a CONTINUUM, not three points, and it makes a prediction
|
|
that can be checked without choosing anything: expR must fall monotonically as the order is
|
|
placed deeper. It also says where to look for the near-zero base the context modifier needs -
|
|
on the SHALLOW side, past the price edge, where nobody in either book places an order.
|
|
|
|
phi < 1 a shallow pullback that never reaches the broken edge - nobody trades here
|
|
phi = 1 the price edge itself - the classic Last Point of Support
|
|
phi > 1 through the edge into the old range: value-area edge and VPOC territory
|
|
|
|
ARMS
|
|
----
|
|
real limit at price_now - phi*(price_now - edge)
|
|
placebo limit at the same DISTANCE from the same starting price, distances permuted
|
|
across events. Geometry identical, level identity destroyed. Without it, a
|
|
shallow-side profit is indistinguishable from "buying small dips works".
|
|
|
|
The placebo is the arm that matters most. A leak or a drift effect lifts BOTH curves; only a
|
|
gap between them is about the level.
|
|
"""
|
|
import numpy as np, sys, time
|
|
sys.stdout.reconfigure(encoding='utf-8', errors='replace')
|
|
import fills, book, wyckoff
|
|
|
|
SYMS = ('EURUSD', 'USDJPY', 'XAUUSD', 'SP500')
|
|
DEPTHS = (0.25, 0.5, 0.75, 1.0, 1.25, 1.5, 2.0)
|
|
|
|
|
|
def curve(sym, tf, kR=2.0, mrisk=2.0, wait=40, H=200, depths=DEPTHS, seed=11):
|
|
bk = fills.Book(sym)
|
|
f = book.frame(sym, tf, bk)
|
|
ev = wyckoff.breakouts(f)
|
|
if ev is None:
|
|
return None
|
|
rows = []
|
|
for u in depths:
|
|
a = wyckoff.retest(sym, tf, u, mrisk=mrisk, kR=kR, wait=wait, H=H,
|
|
bk=bk, f=f, ev=ev)
|
|
b = wyckoff.retest(sym, tf, u, mrisk=mrisk, kR=kR, wait=wait, H=H,
|
|
bk=bk, f=f, ev=ev, placebo=seed)
|
|
rows.append((u, a, b))
|
|
return rows, ev
|
|
|
|
|
|
def show(sym, tf, rows):
|
|
print(f"\n --- {sym} {tf} ---")
|
|
print(f" {'phi':>6}{'n':>7}{'ind':>7}{'fill%':>7}{'REAL':>9}{'t':>7}"
|
|
f"{'placebo':>9}{'t':>7}{'real-plac':>11}{'unres%':>8}")
|
|
us, re, pl = [], [], []
|
|
for u, a, b in rows:
|
|
if a is None:
|
|
continue
|
|
A = a['R'][a['indep']]
|
|
B = b['R'][b['indep']] if b is not None else np.array([0.0])
|
|
us.append(u); re.append(A.mean()); pl.append(B.mean())
|
|
print(f" {u:>+6.2f}{a['n']:>7}{int(a['indep'].sum()):>7}"
|
|
f"{100*a['n']/max(a['placed'],1):>6.1f}%{A.mean():>+9.4f}{book.tstat(A):>+7.2f}"
|
|
f"{B.mean():>+9.4f}{book.tstat(B):>+7.2f}"
|
|
f"{A.mean()-B.mean():>+11.4f}{100*a['unresolved']:>7.1f}%")
|
|
if len(us) >= 4:
|
|
s1, t1 = book.slope_t(np.array(re), np.array(us))
|
|
s2, t2 = book.slope_t(np.array(pl), np.array(us))
|
|
print(f" slope vs depth: real {s1:+.4f} (t {t1:+.2f}) "
|
|
f"placebo {s2:+.4f} (t {t2:+.2f})")
|
|
return us, re, pl
|
|
|
|
|
|
if __name__ == '__main__':
|
|
syms = [s for s in sys.argv[1:] if s in SYMS] or list(SYMS)
|
|
print("=== RETEST DEPTH CURVE ===")
|
|
print(" prediction from the adverse-selection mechanism: expR falls as phi RISES.")
|
|
print(" phi<1 is the shallow side nobody trades - where a near-zero base could live.")
|
|
allrows = []
|
|
for sym in syms:
|
|
for tf in ('H1', 'H4'):
|
|
t0 = time.time()
|
|
out = curve(sym, tf)
|
|
if out is None:
|
|
print(f"\n --- {sym} {tf} --- no events"); continue
|
|
rows, ev = out
|
|
us, re, pl = show(sym, tf, rows)
|
|
print(f" ({len(ev['i']):,} breakouts, {time.time()-t0:.0f}s)")
|
|
allrows.append((sym, tf, us, re, pl))
|
|
if allrows:
|
|
print("\n=== POOLED SHAPE ===")
|
|
print(f" {'phi':>6}{'mean real':>11}{'mean placebo':>14}{'cells real>plac':>17}")
|
|
for k, u in enumerate(DEPTHS):
|
|
r = [re[k] for _, _, us, re, pl in allrows if k < len(re)]
|
|
p = [pl[k] for _, _, us, re, pl in allrows if k < len(pl)]
|
|
if not r:
|
|
continue
|
|
w = sum(1 for x, y in zip(r, p) if x > y)
|
|
print(f" {u:>+6.2f}{np.mean(r):>+11.4f}{np.mean(p):>+14.4f}{w:>10}/{len(r)}")
|