mohan170802/nse-nifty50-swing-predictor

NSE Nifty 50 Swing Trading Predictor v9 — PRODUCTION UNIFIED MODEL

Author: mohan170802 Models: mohan170802/nse-nifty50-swing-predictor

Single Unified Model: 60.3% Test Accuracy on Unseen Data

After extensive data science and ML engineering research, we achieved 60.3% test accuracy with a single unified model that predicts BUY vs NOT_BUY for all 50 NSE Nifty stocks simultaneously.

Architecture: Why This Works

Previous Approach (v7): 7 Sector Models

• —One model per sector: BANKING, IT, AUTO, ENERGY, FMCG, PHARMA, DIVERSIFIED
• —Mean accuracy: 57.7% | Best: IT 62.5%
• —Problem: Requires maintaining 7 models, no cross-sector learning

v9: Single Unified Cross-Sectional Model ✅

• —One CatBoost model trained on ALL 50 stocks together
• —Cross-sectional features: z-scores and percentile ranks computed within each trading date across all stocks
• —Temporal target encoding: Expanding mean of BUY rate per ticker/sector (no leakage)
• —Time-decay weighting: Recent samples weighted more heavily
• —Native categorical handling: ticker (50 levels) and sector (7 levels) as CatBoost categorical features

Results

Winner: CatBoost — better categorical feature handling, ordered target statistics prevent leakage

Top 15 Features (by importance)

Key insight: vix_level and month dominate — market regime and seasonality matter more than any single stock indicator. Cross-sectional relative features (*_rel_z) contribute significantly.

Feature Engineering (105 features)

Technical Indicators

• —Momentum: RSI(14), RSI(28), MACD, Stochastic, ADX
• —Volatility: Bollinger Bands %B & width, ATR normalized
• —Volume: OBV, VWAP deviation, volume ratio

Returns & Trend

• —Lagged log returns: 1, 2, 3, 5, 10 days
• —Rolling: mean/volatility 5, 10, 20 days
• —Price action: HL range, distance from SMA(5/10/20/50), trend 5/20, momentum 10/20

Cross-Sectional (Relative) — THE KEY INNOVATION

• —For every raw feature: feature_rel_z = z-score across all 50 stocks on same date
• —For every raw feature: feature_rel_pct = percentile rank across all 50 stocks
• —Sector-relative: z-score within each sector on same date
• —Market percentile: stock's return rank vs all stocks

Macro & Calendar

• —Nifty50 return, India VIX level & 5d change, relative to Nifty
• —Day-of-week, month, ticker, sector

Temporal Target Encoding (No Leakage)

• —ticker_target_enc: Expanding mean of BUY rate per ticker, shifted by 1
• —sector_target_enc: Expanding mean of BUY rate per sector, shifted by 1

Training Details

Production Usage

Quick Predict

python nse_v9_predict.py --ticker RELIANCE.NS --threshold 0.55

Python API

import catboost as cb
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download("mohan170802/nse-nifty50-swing-predictor", "unified_model_v9.cbm")
model = cb.CatBoostClassifier()
model.load_model(model_path)

# After engineering same 105 features from latest data:
# proba = model.predict_proba(features)[0, 1]
# signal = "BUY" if proba > 0.55 else "NOT_BUY"

Thresholds for Swing Trading

Files

unified_model_v9.cbm          # CatBoost binary model
unified_model_v9.json         # XGBoost backup model
unified_features_v9.txt       # 105 feature names (load order critical)
label_encoders_v9.json        # Ticker/sector label mappings
summary_unified_v9.json       # Full metrics, top features, model comparison
nse_v9_predict.py             # Production inference script
nse_v9_final.py               # Training script (reproducible)

Performance Reality Check

• —Random baseline: 50% (binary)
• —Our model: 60.3% on completely unseen test data
• —Edge: +10.3 percentage points above random
• —AUC: 0.559 (weak but consistent signal)
• —F1: 0.337 (low due to class imbalance — 43% buy rate)

Interpretation: This is a weak but real signal. In trading:

• —60% directional accuracy with proper position sizing and stop-losses can be profitable
• —Use as ONE signal among many (not standalone)
• —The model identifies when market regime + stock setup align for above-average probability of +3% move

Improvements for v10+

1. 1.Fundamental features: P/E, ROE, EPS growth, debt/equity (from NSE/Indian exchanges)
2. 2.News/sentiment: Earnings announcements, macro events
3. 3.Options data: Implied volatility, put/call ratio
4. 4.More history: 10+ years to capture more market cycles
5. 5.Quarterly retraining: As recommended by Yang et al. research
6. 6.Ensemble with deep learning: Transformer models (MASTER paper) for cross-sectional attention

⚠️ Disclaimer

Not financial advice. Past performance does not guarantee future results. Use with:

• —Stop-losses (suggest -2%)
• —Position sizing (risk ≤2% per trade)
• —Portfolio-level risk management
• —Paper trading before real capital

Citation

Built using insights from:

• —CatBoost: unbiased boosting with categorical features (Dorogush et al., NeurIPS 2017)
• —MASTER: Market-Guided Stock Transformer (Li et al., KDD 2023)
• —Marcos López de Prado: Advances in Financial Machine Learning (Triple Barrier Method)

<!-- ml-intern-provenance -->

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 = 'mohan170802/nse-nifty50-swing-predictor'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

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