27 lines
1.7 KiB
Text
27 lines
1.7 KiB
Text
=== POST-PURGE BASELINE (SB-01 + SB-02 fixed) ===
|
|
Tanggal : 2026-08-21
|
|
Command : python train_model.py XAUUSDc --out repaired\post_purge\SniperGold_ML_postpurge.mqh --tag v20260821_postpurge
|
|
Hash kode baru (train_model.py, setelah edit) : (lihat provenance hash di bawah)
|
|
Dataset (features_XAUUSDc.npz) : E60285F55AB5AF7B397653373F0EE5B8B814708DC178DD58691DAE0BE122946A
|
|
Model POST : repaired\post_purge\SniperGold_ML_postpurge.mqh (SHA256 9D8906A89DB6580A01A1F1747E6E24D4C9D3F849910AEDCE2AFE873B6F1FE44E)
|
|
Determinisme : dijalankan 2x -> hasil IDENTIK (reproducible).
|
|
|
|
Hasil (kode baru; SB-01 metrik validasi per-head; SB-02 purge gap=H_LABEL):
|
|
SB-02 purged split: split_bar=44972 train=35652 purge=13 test=11872
|
|
max_train_outcome_end=44995 < min_test_outcome_start=44997 OK
|
|
AUC validasi (SB-01 corrected): LONG=0.6796 SHORT=0.6796 mean=0.6796 (epoch terbaik 55)
|
|
LONG : TRAIN AUC 0.7483 | VAL 0.6796 | TEST 0.6582
|
|
SHORT: TEST AUC 0.6582
|
|
CALIBRASI LONG A=0.6913 B=0.1935 | SHORT A=0.6913 B=-0.1935
|
|
n_tr=30304 n_purge=13 n_te=11872 split_bar=44972 tag=v20260821_postpurge
|
|
|
|
PERBANDINGAN PRE vs POST:
|
|
PRE (kode lama, metrik rusak, split tanpa purge): TEST 0.6270 / 0.6207 ; best_epoch=0
|
|
POST (kode baru, metrik benar, purge gap=24) : TEST 0.6582 / 0.6582 ; best_epoch=55
|
|
Interpretasi: KOREKSI METODOLOGI (bukan optimasi). Freeze lama = model ~1 epoch
|
|
(early stopping memilih epoch 0 karena metrik validasi rusak ~0.5 konstan).
|
|
|
|
CATATAN:
|
|
- angka ini adalah ESTIMASI terkoreksi pada cache features_XAUUSDc.npz (feed XAUUSDc,
|
|
split 75/25 dengan purge). Bukan klaim edge; belum divalidasi lintas feed.
|
|
- freeze Include\SniperGold_ML.mqh TIDAK diubah.
|