Kronu/nq-futures-path-a

NQ Futures Key Level Classifier

LightGBM classifier for predicting NQ futures key level reactions based on market microstructure

Model Details

• Model Type: LightGBM Classifier
• Task: Multi-class classification for NQ futures key level reactions
• Training Data: 1,043,089 episodes from 78 trading days (April-August 2025)
• Features: 0 market microstructure and session context features

Performance

• Multi-class Log Loss: 0.9852353681288883
• Accuracy: 0.9764018445196484

Feature Importance (Top 10)

• mae_ticks_60s: 6659.0000
• rv_60s: 6409.0000
• mfe_ticks_60s: 6190.0000
• touch_count_last_30m: 5442.0000
• vwap_dev_ticks: 5023.0000
• mfe_ticks_120s: 4985.0000
• session_cum_delta: 4864.0000
• mae_ticks_120s: 4795.0000
• pullback_ticks_30s: 4774.0000
• p50_intertrade_ms_5s: 4770.0000

Usage

import joblib
import pandas as pd

# Load the model
model = joblib.load('classifier.joblib')

# Prepare features (same format as training)
features = prepare_features(your_data)

# Make predictions
predictions = model.predict(features)
probabilities = model.predict_proba(features)

Model Architecture

This model predicts four possible outcomes when price approaches key levels:

• BREAK: Price breaks through the level decisively
• BOUNCE: Price bounces off the level
• WEAK_BREAK: Price breaks but with weak momentum
• TIMEOUT: Price approaches but doesn't reach outcome within time limit

Training Context

The model was trained on NQ futures data from the first 2 hours of regular trading hours (09:30-11:30 ET), focusing on:

• Key level identification (OPEN, IBH/IBL, Round Numbers, Session VWAP)
• Market microstructure features (order flow, volatility, timing)
• Session context (cumulative delta, VWAP deviation, touch frequency)

License

MIT License - see LICENSE file for details.