56 lines
2.3 KiB
Python
56 lines
2.3 KiB
Python
# -*- coding: utf-8 -*-
|
|
"""Verifikasi akhir features_XAUUSD.npz: shape, NaN, variasi fitur, densitas label."""
|
|
import os
|
|
import json
|
|
import numpy as np
|
|
import datetime as dt
|
|
|
|
BASE = os.path.normpath(os.path.join(os.path.dirname(os.path.abspath(__file__)),
|
|
"..", "..", "..", "Files", "AlgoForge", "Data"))
|
|
z = np.load(os.path.join(BASE, "features_XAUUSD.npz"))
|
|
F, F2, label, ATR, close, t = (z["F"], z["F2"], z["label"], z["ATR"],
|
|
z["close"], z["time"])
|
|
split_bar, split_pos = int(z["split_bar"]), int(z["split_pos"])
|
|
meta = json.loads(str(z["meta"]))
|
|
print(f"F={F.shape} F2={F2.shape} label={label.shape} ATR={ATR.shape} "
|
|
f"close={close.shape} time={t.shape}")
|
|
print(f"split_bar={split_bar} split_pos={split_pos}")
|
|
|
|
for nm, X in (("F", F), ("F2", F2)):
|
|
n_nan = int(np.isnan(X).sum()) + int(np.isinf(X).sum())
|
|
stds = X.std(0)
|
|
n_const = int((stds < 1e-12).sum())
|
|
print(f"{nm}: NaN/Inf={n_nan} | kolom konstan={n_const} | "
|
|
f"std min={stds.min():.3e} max={stds.max():.3e}")
|
|
|
|
# fitur F2 terakhir (regime) rentang
|
|
print("F2 kolom terakhir (p_lowvol/p_highvol): min=%.3f max=%.3f"
|
|
% (F2[:, 4].min(), F2[:, 5].max()))
|
|
|
|
# label & window
|
|
win = t >= dt.datetime(2018, 1, 1).timestamp()
|
|
mask = (label != 0) & win
|
|
midx = np.where(mask)[0]
|
|
n_bull = int((label[midx] == 1).sum())
|
|
n_bear = int((label[midx] == -1).sum())
|
|
tr, te = midx[:split_pos], midx[split_pos:]
|
|
print(f"\nberlabel={len(midx)} bull={n_bull} bear={n_bear} "
|
|
f"ratio={n_bull/n_bear:.3f}")
|
|
print(f"train={len(tr)} (bull={(label[tr]==1).mean():.3f}) | "
|
|
f"test={len(te)} (bull={(label[te]==1).mean():.3f})")
|
|
|
|
# densitas label: 24-bar forward = berapa jam sesi?
|
|
dts = t[midx + 24] - t[midx]
|
|
hours = dts / 3600.0
|
|
print(f"\n24-bar forward: median={np.median(hours):.2f} jam | "
|
|
f"p10={np.percentile(hours, 10):.2f} p90={np.percentile(hours, 90):.2f} "
|
|
f"| >8 jam: {(hours > 8).mean()*100:.1f}%")
|
|
|
|
# cek feature F untuk beberapa bar acak (sanity numerik)
|
|
i = midx[100000]
|
|
print(f"\ncontoh bar t={dt.datetime.fromtimestamp(int(t[i]), dt.timezone.utc)}: "
|
|
f"label={label[i]} ATR={ATR[i]:.4f} close={close[i]:.2f}")
|
|
print("F[:5] =", np.round(F[i, :5], 3))
|
|
print("F[15:19] =", np.round(F[i, 15:], 3))
|
|
print("F2 =", np.round(F2[i], 3))
|
|
print("\nOK.")
|