AcyLa/forex-predictive-model

🏦 Forex Predictive Model — ML Ensemble with Nadaraya-Watson + Bollinger

A production-grade machine learning pipeline for intraday forex price prediction using Nadaraya-Watson envelope estimation, multi-window Bollinger Bands, and 180+ technical indicators. Outputs BUY / SELL / NEUTRAL signals with pip-based price targets.

🚀 Quick Start

pip install yfinance pandas numpy scikit-learn lightgbm ta torch pyarrow joblib scipy
git clone https://huggingface.co/AcyLa/forex-predictive-model
cd forex-predictive-model

# Get predictions immediately (pre-trained models included)
python predict.py --ticker EURUSD=X

# Fast mode (LightGBM only, sub-ms)
python predict.py --ticker EURUSD=X --no-lstm

# Single horizon
python predict.py --ticker EURUSD=X --horizon 8 --features

🏗 Architecture

Yahoo Finance FX Data → 180+ Features → Multi-Horizon Labels → LightGBM + LSTM Ensemble → BUY/SELL/NEUTRAL
         ↑                    ↑                                                                          ↑
   730 days 1h bars    Nadaraya-Watson                                                     Pip-based targets
                       Bollinger (3 windows)                                               Consensus voting
                       RSI, MACD, ADX, Ichimoku
                       Stochastic, ATR, Donchian
                       Session timing (London/NY/Tokyo)

📊 Key Components

Nadaraya-Watson Envelope (12 features)

Non-parametric kernel regression that smooths price to identify trend + overbought/oversold zones. Vectorized implementation computes 17k bars in 0.3s.

Bollinger Bands System (45 features)

• 3 windows (20, 50, 100) × 3 std devs (2.0, 2.5, 3.0)
• Bandwidth, %B position, squeeze detection, band-walk signals

All Best Indicators (180+ total)

Category	Indicators
Trend	EMA stack (9/21/50/100/200), SMA, MACD, ADX, Ichimoku Cloud, CCI
Momentum	RSI (7/14/21), Stochastic, Williams %R, ROC, Awesome Oscillator
Volatility	ATR (14/21/50), Bollinger, Keltner, Donchian, Historical Vol
Volume	OBV, MFI, VWAP, CMF, Volume ratios
Price	Multi-horizon returns, candlestick, support/resistance, drawdown
Session	Tokyo/London/NY sessions, overlaps, opens/closes, day-of-week
NW	Smooth, envelopes, slope, crossovers, trend regime, breakout signals

Trading Session Awareness

The model knows when London opens (08:00 UTC), NY opens (13:00 UTC), and the high-activity London-NY overlap (13:00-17:00 UTC). This is critical for FX — most price action happens during session overlaps.

📈 Prediction Output

══════════════════════════════════════════════════════
  EURUSD=X | Multi-Horizon FX Predictions
  Latest Bar:    2026-09-11 22:00
  Current Price: 1.16023
══════════════════════════════════════════════════════
   1h: 🟡 NEUTRAL  |   -0.0 pips | target=1.16023
   4h: 🟡 NEUTRAL  |   +0.2 pips | target=1.16025
   8h: 🟡 NEUTRAL  |   +0.8 pips | target=1.16030
  24h: 🟡 NEUTRAL  |   +3.9 pips | target=1.16062
──────────────────────────────────────────────────────
  CONSENSUS: 🟡 NEUTRAL  (confidence: 50%)
══════════════════════════════════════════════════════

🔧 Python API

from predict import ForexPredictor

predictor = ForexPredictor(ticker="EURUSD=X")

# Multi-horizon with consensus
result = predictor.get_multi_horizon()
print(f"Consensus: {result['consensus']}")

# Single horizon
signal = predictor.get_signal(horizon=8)
# signal = {
#   'pair': 'EURUSD=X', 'signal': 'BUY', 'confidence': 0.65,
#   'predicted_pips': 12.3, 'current_price': 1.16023,
#   'target_price': 1.16146, 'horizon_hours': 8
# }

# Fast HFT mode (LightGBM only)
fast = predictor.get_signal(horizon=1, use_lstm=False)

# Batch prediction on your own data
preds = predictor.predict_batch(your_ohlcv_df, horizon=4)

🔄 Training Other Pairs

# Train on GBP/USD
python train.py --ticker GBPUSD=X --no-backtest

# Train on USD/JPY (pip size auto-detected as 0.01)
python train.py --ticker USDJPY=X --no-backtest

# Full walk-forward backtest (slower)
python train.py --ticker EURUSD=X

Supported pairs: EURUSD, GBPUSD, USDJPY, USDCHF, AUDUSD, USDCAD, NZDUSD

📁 Files

forex-predictive-model/
├── config.py              # All hyperparameters
├── data_loader.py         # Yahoo Finance FX download
├── nadaraya_watson.py     # NW envelope estimator (vectorized)
├── features.py            # 180+ indicators (Bollinger, RSI, MACD, etc.)
├── labels.py              # Pip-based labels + triple-barrier
├── models.py              # LightGBM + LSTM ensemble
├── backtest.py            # Walk-forward backtester (pip P&L)
├── train.py               # Training pipeline CLI
├── predict.py             # Prediction API + CLI
└── fx_models/EURUSD_X/   # Pre-trained models (4 horizons)

⚠️ Disclaimer

Research tool, not financial advice. FX trading carries high risk. Always use stop-losses and proper position sizing.

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = 'AcyLa/forex-predictive-model'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.