LGB baseline model card

checkpoints/lgb_baseline/MODEL_CARD.md

@@ -0,0 +1,46 @@
+# LightGBM Baseline — BTC/USDT 15-min Direction
+
+## What it does
+Classifies whether BTC/USDT will rise >0.05% or fall >0.05% over the next 15 min
+using **119 features** (base + lags + rolling stats).
+
+Output: probability P(up) ∈ [0, 1].
+
+## When to use
+- Fast baseline or fallback when transformer is unavailable
+- Feature importance analysis (gain-based ranking)
+- Ensemble component alongside CryptoTransformer
+- **Instant** inference (< 1ms per bar)
+- **No GPU required**
+
+## Performance
+| Metric | Value |
+|--------|-------|
+| Val AUC | **0.5476** |
+| Val log-loss | 0.6899 |
+| Best iteration | 48 |
+| Val fraction | last 15% of data |
+| Features | 119 (base + lags t-1/3/6/12/24 + rolling 30m/1h/2h/4h) |
+
+## Feature engineering
+- **Base**: 20 raw 5-min features (see feature schema)
+- **Lags**: t-1, t-3, t-6, t-12, t-24 bars for 9 key microstructure signals
+- **Rolling**: mean + std over 6, 12, 24, 48 bar windows for 6 signals
+- **Time**: hour_utc, sin/cos(hour), day_of_week, sin/cos(dow)
+
+## Top features (by GBDT gain)
+`depth_imbalance_1pct`, `log_ret_15m`, `rsi_14`, `log_ret_60m`,
+`taker_buy_ratio_5m_rmean6`, `oi_btc`, `vpin_50_rstd48`
+
+## How to load
+```python
+import lightgbm as lgb
+import json
+
+model    = lgb.Booster(model_file="lgb_model.txt")
+meta     = json.load(open("meta.json"))
+feat_cols = meta["feat_cols"]
+
+# Build feature row (single-row numpy array in feat_cols order)
+prob_up = model.predict(X_row)[0]   # X_row shape (1, n_features)
+```