PREDICTIVE_ML.md

Predictive ML (frozen)

Next-candle direction classification on 1m klines. Kept for reference and possible use as a regime filter — not a standalone trading strategy.

Status: Frozen — accuracy does not survive realistic fees at 1m horizons.

Install

pip install -e ".[prediction]"

Layout

src/strategy/predictive/
├── features.py   # 80+ technical features from OHLCV
├── models.py     # Logistic, RF, XGBoost, GBM, MLP, CNN, LSTM, Transformer
└── train.py      # ModelTrainer, TimeSeriesSplitter, FeatureSelector

Minimal example

from src.data.data_load import load_data
from src.strategy.predictive import (
    prepare_data_for_modeling,
    TimeSeriesSplitter,
    RandomForestModel,
)

df = load_data(symbols="BTCUSDT", start_date="2025-01-01", end_date="2025-06-30", interval="1m")
df_features, feature_cols, target_col = prepare_data_for_modeling(df)

X = df_features[feature_cols].values
y = df_features[target_col].values
X_train, X_val, X_test, y_train, y_val, y_test = TimeSeriesSplitter.train_val_test_split(X, y)

model = RandomForestModel(n_estimators=100)
model.train(X_train, y_train)
print(f"Test accuracy: {(model.predict(X_test) == y_test).mean():.2%}")

Runnable examples

• Script: scripts/examples_price_prediction.py
• Notebook: notebooks/price_forecast_complete.ipynb

Models

Model	Use case
LogisticRegression	Fast baseline
RandomForest / XGBoost / GBM	Best tabular accuracy
MLP / CNN / LSTM / Transformer	Sequence experiments (slow, needs GPU for scale)

Always use TimeSeriesSplitter — never shuffle time series.

If revisiting

• Predict at 15m–1h horizons, not 1m
• Use model output to filter classical signals (meta-labeling), not direct orders
• Compare against dual_ma benchmark in STRATEGY_STATUS.md