Arkm20/sniper-model-v7

🎯 Sniper Model v7.1 — Asymmetric Precision Hunter

A LightGBM-based stock signal model optimized for high-precision buy signals using a dual-barrier labeling methodology and focal loss training.

Model Description

Sniper Model v7.1 is a tabular classification model that predicts high-probability stock price breakouts over a 15-day horizon. It is trained on 800+ US equities (2005–2018) with walk-forward cross-validation and Optuna hyperparameter optimization.

Key Features

• 100+ engineered technical features across 6 categories (exhaustion, oscillators, volatility, trend, lagged, external)
• Dual-barrier labeling: Profit target = 3× ATR, Stop loss = 0.5× ATR
• Focal loss (α=0.25, γ=2.0) for hard-example mining
• Two-stage funnel: Stage 1 filters top 25%, Stage 2 refines with focal loss
• Regime ensemble: Separate models for bull/bear × low/high VIX environments
• Isotonic calibration for reliable probability estimates

Architecture

• Base model: LightGBM with Optuna-tuned hyperparameters (150 trials)
• Walk-forward evaluation: 8 outer folds with 20-day purge gaps
• Feature selection: LightGBM importance → SHAP filtering

Training Data

• Universe: ~800 US equities (blended S&P 500 + Dow + Russell 2000)
• Training period: 2005-01-01 to 2018-12-31
• Holdout period: 2019-01-01 to 2020-12-31

Usage

import joblib
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download("Arkm20/sniper-model-v7", "models/lgb_model_*.pkl")
model = joblib.load(model_path)

# Predict
probabilities = model.predict_proba(features)[:, 1]

Associated Space

• Live Scanner: Arkm20/stock-scanner

Limitations & Disclaimer

• Trained on historical data (2005–2020); market regimes change
• NOT financial advice — educational/research purposes only
• Past performance does not guarantee future results