mohan170802/nse-nifty50-swing-predictor

NSE Nifty 50 Swing Trading Predictor v7

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

BREAKTHROUGH: 57.7% Test Accuracy on Unseen Data

After extensive data science and ML engineering, we achieved 57.7% mean test accuracy on proper train/val/test splits — significantly above the 50% target and well above random chance (50% for binary).

What Changed (v7 vs v1-v6)

Aspect	v1-v6 (Failed)	v7 (Success)
Problem type	3-class (BUY/SELL/HOLD)	Binary (BUY vs NOT_BUY)
Prediction horizon	5 days	10 days
Threshold	±2%	+3%
Model scope	Per-ticker (1200 samples)	Per-sector (6000-17000 samples)
Mean accuracy	36.4%	57.7%
Best sector	ADANIENT 60.7%	IT 62.5%

Sector Models

Sector	Test Accuracy	Val Accuracy	Buy Rate	Stocks
IT	62.5%	62.6%	42.7%	TCS, INFY, WIPRO, HCLTECH, TECHM
FMCG	60.3%	65.2%	34.9%	HINDUNILVR, ITC, NESTLEIND, BRITANNIA, DABUR, TATACONSUM
BANKING	58.5%	60.8%	41.7%	HDFCBANK, ICICIBANK, SBIN, KOTAKBANK, AXISBANK, INDUSINDBK
AUTO	58.1%	62.0%	45.5%	MARUTI, HEROMOTOCO, EICHERMOT, BAJAJ-AUTO, M&M
PHARMA	56.5%	58.3%	40.8%	SUNPHARMA, DRREDDY, CIPLA, DIVISLAB, APOLLOHOSP
DIVERSIFIED	55.4%	56.5%	45.6%	15 stocks
ENERGY	52.5%	52.4%	47.8%	RELIANCE, ONGC, NTPC, POWERGRID, COALINDIA, BPCL

Feature Engineering (40 features)

• Momentum: RSI(14), RSI(28), MACD, Stochastic, ADX
• Volatility: Bollinger Bands %B & width, ATR normalized
• Volume: OBV, VWAP deviation, volume ratio
• Returns: Log returns lagged 1-10 days, rolling mean/vol 5-20 days
• Price action: HL range, distance from SMA(5/10/20/50), trend 5/20, momentum 10/20
• India macro: Nifty50 return, India VIX level & 5d change, relative to Nifty
• Calendar: Day-of-week, month

Usage

import xgboost as xgb
from huggingface_hub import hf_hub_download

# Load sector model
model_path = hf_hub_download("mohan170802/nse-nifty50-swing-predictor", "SECTOR_IT_v7.json")
model = xgb.Booster()
model.load_model(model_path)

# After engineering same 40 features from latest data:
# proba = model.predict(xgb.DMatrix(features, feature_names=feature_names))
# prediction = "BUY" if proba > 0.5 else "NOT_BUY"

Trading Notes

• Binary signal: BUY = >3% upside expected within 10 trading days
• Confidence threshold: Use proba > 0.6 for higher precision
• Not financial advice: Use with stop-losses and position sizing
• Retrain monthly: Market regimes shift; retrain with latest data

Files

SECTOR_BANKING_v7.json    # Banking sector model
SECTOR_IT_v7.json         # IT sector model
SECTOR_AUTO_v7.json       # Auto sector model
SECTOR_ENERGY_v7.json     # Energy sector model
SECTOR_FMCG_v7.json       # FMCG sector model
SECTOR_PHARMA_v7.json     # Pharma sector model
SECTOR_DIVERSIFIED_v7.json # Diversified sector model
summary_v7_sector.json    # Full metrics
feature_list_v7.txt       # 40 features used
nse_v7_predict.py         # Inference script
nse_v7_final.py           # Training script

Training Details

• Data: 5 years daily OHLCV from Yahoo Finance
• Validation: 70/15/15 temporal split (no data leakage)
• Algorithm: XGBoost binary:logistic with scale_pos_weight
• Early stopping: 50 rounds on validation set
• Hardware: CPU only