XauBot/docs/arsitektur-ai/02-XGBoost-Signal-Predictor.md

XGBoost — Signal Predictor

File: src/ml_model.py Model: models/xgboost_model.pkl Library: xgboost

---

Apa Itu XGBoost?

XGBoost (eXtreme Gradient Boosting) adalah algoritma machine learning berbasis ensemble decision tree. Model ini belajar dari puluhan fitur teknikal untuk memprediksi arah harga di bar berikutnya.

Analogi: XGBoost adalah navigator AI — menentukan apakah harga akan naik atau turun.

---

Fungsi Utama

XGBoost bertugas memprediksi probabilitas harga naik atau turun di bar M15 berikutnya, lalu menghasilkan signal BUY, SELL, atau HOLD.

prob_up > 0.65   ->  BUY
prob_down > 0.65 ->  SELL
lainnya          ->  HOLD (tidak cukup yakin)

---

Arsitektur Model

params = {
    "objective": "binary:logistic",   # Klasifikasi biner (naik/turun)
    "eval_metric": "auc",             # Area Under Curve
    "max_depth": 3,                   # Kedalaman tree (anti-overfitting)
    "learning_rate": 0.05,            # Lambat & stabil
    "min_child_weight": 10,           # Minimum sampel per leaf
    "subsample": 0.7,                 # 70% data per round
    "colsample_bytree": 0.6,          # 60% fitur per tree
    "reg_alpha": 1.0,                 # L1 regularization
    "reg_lambda": 5.0,                # L2 regularization (kuat)
    "gamma": 1.0,                     # Min loss reduction per split
}

Anti-Overfitting:

• Tree dangkal (depth 3, bukan 6)
• Early stopping setelah 5 round tanpa improvement
• Feature subsampling 60%
• Regularisasi L2 kuat (lambda=5.0)

---

Input (24 Fitur)

Indikator Teknikal

Fitur	Sumber	Fungsi
rsi	Feature Eng	Overbought/oversold
atr, atr_percent	Feature Eng	Volatilitas
macd, macd_signal, macd_histogram	Feature Eng	Momentum tren
bb_percent_b, bb_width	Feature Eng	Posisi dalam Bollinger Band
ema_9, ema_21	Feature Eng	Tren jangka pendek

Returns & Momentum

Fitur	Formula	Fungsi
returns_1	close[t]/close[t-1] - 1	Return 1 bar
returns_5	close[t]/close[t-5] - 1	Return 5 bar
returns_20	close[t]/close[t-20] - 1	Return 20 bar
log_returns	ln(close[t]/close[t-1])	Log return

Volatilitas & Posisi Harga

Fitur	Fungsi
volatility_20	Realized volatility 20 bar
normalized_range	(High-Low)/Close
avg_normalized_range	Rata-rata range 14 bar
price_position	Posisi 0-1 dalam range
dist_from_sma_20	Jarak dari SMA 20

Smart Money Concepts (SMC)

Fitur	Fungsi
swing_high, swing_low	Fractal structure
fvg_signal	Fair Value Gap (1/-1/0)
ob	Order Block (1/-1/0)
bos, choch	Break of Structure, Change of Character
market_structure	Bullish/Bearish (1/-1/0)

Waktu & Regime

Fitur	Fungsi
hour, weekday	Pola jam & hari
london_session, ny_session	Flag sesi trading
regime	HMM regime state (0/1/2)

---

Cara Kerja

Proses Prediksi (Setiap Loop)

DataFrame lengkap (200 bar + semua fitur)
        |
        v
Ambil baris terakhir (1 bar)
        |
        v
Pilih 24 fitur yang sesuai dengan training
        |
        v
Buat DMatrix (format XGBoost)
        |
        v
model.predict() -> probabilitas harga NAIK (0-1)
        |
        v
Tentukan signal:
  prob_up > 0.65   -> BUY
  prob_down > 0.65 -> SELL
  lainnya          -> HOLD
        |
        v
Output: PredictionResult
  - signal: "BUY" / "SELL" / "HOLD"
  - probability: 0-1 (prob naik)
  - confidence: max(prob_up, prob_down)
  - feature_importance: {fitur: skor}

Proses Training

1. Ambil 10,000 bar M15 XAUUSD
2. Feature engineering (40+ kolom)
3. SMC analysis (swing, FVG, OB, BOS, CHoCH)
4. Buat target: 1 jika close[t+1] > close[t], else 0
5. Split: 70% train, 30% test
   PENTING: 50-bar gap antara train & test set (v4)
   → Mencegah temporal leakage (autocorrelation antar bar berdekatan)
   → Train: bar 0 sampai split_point
   → Test: bar split_point + 50 sampai akhir
6. Train XGBoost 50 round + early stopping (patience=5)
7. Evaluasi: Train AUC vs Test AUC
8. Simpan model + feature names ke .pkl

---

Output & Dampak ke Trading

1. Validasi Signal SMC

SMC bilang BUY + XGBoost setuju (>55%)    -> TRADE
SMC bilang BUY + XGBoost netral (<55%)    -> SKIP
SMC bilang BUY + XGBoost bilang SELL >75% -> TOLAK (veto)

2. Confidence Gate

ML confidence < 55%  -> Tidak boleh entry (terlalu tidak yakin)
ML confidence 55-65% -> Entry dengan lot kecil
ML confidence > 65%  -> Entry dengan lot penuh

3. Exit Signal (Penutupan Posisi)

Posisi BUY terbuka
XGBoost prediksi SELL dengan confidence > 75%
-> TUTUP posisi (ML reversal exit)

4. Feature Importance

# Contoh output (top 5)
{
    "market_structure": 0.85,   # Fitur paling penting
    "rsi": 0.68,
    "atr_percent": 0.65,
    "macd_histogram": 0.52,
    "bos": 0.48,
}

Menunjukkan fitur mana yang paling berpengaruh dalam keputusan model.

---

Metrik Evaluasi

{
    "train_auc": 0.6234,      # Performa di data training
    "test_auc": 0.5932,       # Performa di data testing
    "train_samples": 7000,
    "test_samples": 3000,
    "num_features": 24,
}

AUC	Interpretasi
0.50	Sama dengan tebak acak
0.55	Sedikit lebih baik dari acak
< 0.60	ROLLBACK — terlalu rendah untuk trading (v4 threshold)
0.60-0.65	Minimum acceptable, warning
0.65+	Cukup baik untuk trading
0.70+	Sangat baik

Rollback threshold: Jika test AUC < 0.60, model otomatis rollback ke versi sebelumnya. (v4: dinaikkan dari 0.52 karena 0.52 hampir sama dengan tebak acak)

---

Auto-Retraining

• Jadwal: Harian pukul 05:00 WIB
• Data: 8,000 bar (daily) / 15,000 bar (weekend deep training)
• Cek retrain: Setiap 20 candle M15 (~5 jam) — candle-based, bukan time-based
• Proses: Backup lama -> retrain -> validasi AUC -> simpan/rollback
• Rollback: AUC < 0.60 → otomatis rollback (v4: dinaikkan dari 0.52)
• Minimum interval: 20 jam antar retrain (cegah overfitting)
• Train/test gap: 50 bar antara train dan test set (anti temporal leakage)

---

Contoh Skenario

Skenario 1: Signal kuat

RSI=35 (oversold), MACD rising, BOS bullish, market_structure=1
-> XGBoost: prob_up=0.78 -> BUY (confidence 78%)
-> Lot penuh, entry dieksekusi

Skenario 2: Konflik dengan SMC

SMC signal: BUY
XGBoost: prob_down=0.82 -> SELL (confidence 82%)
-> Signal DITOLAK (ML strongly disagrees >75%)
-> Tidak ada trade

Skenario 3: Tidak yakin

RSI=50, MACD flat, regime=1
-> XGBoost: prob_up=0.53 -> HOLD (confidence 53% < 55%)
-> Tidak ada trade — tunggu signal lebih jelas