stocky-ai/stock-predictor-scratch

language: - en - tr tags: - finance - stocks - news-analysis - stock-prediction - information-retrieval license: apache-2.0 datasets: - financial-news metrics: - recall pipeline_tag: zero-shot-classification

📈 Stocky Stock Predictor

Predict stock symbols from news headlines using a two-stage deep learning pipeline.

🎯 Model Overview

This model predicts which stock symbols are mentioned or relevant to a given news headline. It uses a sophisticated two-stage retrieval system:

1. Stage 1 - Contrastive Retrieval: Fast dual-encoder retrieves top-50 candidates
2. Stage 2 - Cross-Encoder Reranking: Precise scoring to get final top-10 predictions

📊 Performance

Test Set Metrics (10,620 samples):

• Recall@5: 51.17%
• Recall@10: 54.40%
• Recall@20: 58.76%

Real-World Performance:

• Direct company mentions (e.g., "Nvidia announces..."): ~80% accuracy
• Generic sector news (e.g., "Tech stocks rally"): ~50% accuracy

🚀 Quick Start

from predict import StockyPredictor

# Initialize predictor
predictor = StockyPredictor()

# Your list of stock symbols (load all 6,973 stocks in production)
stocks = ["NVDA", "AAPL", "MSFT", "GOOGL", "AMZN", "TSLA", "META", ...]

# Predict
title = "Nvidia announces new AI chip breakthrough"
predictions = predictor.predict(title, stocks, top_k=10)

# Results: [('NVDA', 0.95), ('AMD', 0.78), ...]
for stock, score in predictions:
    print(f"{stock}: {score:.4f}")

💡 Use Cases

• News Analysis: Automatically tag news articles with relevant stocks
• Trading Signals: Identify stocks mentioned in breaking news
• Portfolio Monitoring: Track news about your holdings
• Market Research: Analyze media coverage of stocks

🏗️ Model Architecture

Components:

1. Tokenizer (35K vocab)

   • Custom WordPiece trained on financial text
   • Top 500 stocks as special tokens
   • 98.4% coverage of frequent stocks

2. MLM Pretrained Encoder (113M params)

   • BERT-base architecture
   • Pretrained from scratch on FinCorpus + news
   • Perplexity: 5.08

3. Contrastive Model (226M params)

   • Dual-encoder (title encoder + stock encoder)
   • Trained with InfoNCE loss
   • Retrieves top-50 candidates in milliseconds

4. Cross-Encoder Reranker (113M params)

   • BERT + classification head
   • Binary relevance scoring
   • Validation loss: 0.1619

Training Details:

• Dataset: 106,207 news articles, 6,973 unique stocks
• Training time: ~4-5 days on V100 GPU
• Stages: MLM (3 epochs) → Contrastive (5 epochs) → Reranking (3 epochs)

📝 Examples

Example 1: Tech Company

title = "Apple unveils new iPhone with advanced camera"
predictions = predictor.predict(title, all_stocks)
# Top result: AAPL (0.94)

Example 2: Cryptocurrency

title = "Bitcoin price surges past $50,000"
predictions = predictor.predict(title, all_stocks)
# Top result: BTC-USD (0.92)

Example 3: Multiple Companies

title = "Tech giants report strong earnings amid AI boom"
predictions = predictor.predict(title, all_stocks)
# Top results: NVDA (0.89), MSFT (0.85), GOOGL (0.78)

⚠️ Limitations

• Works best when company names are mentioned directly
• Generic market news (e.g., "tech sector rallies") may have lower accuracy
• ETF symbols (SPY, QQQ) have limited training data
• Primarily trained on English news (some Turkish support)

🔧 Advanced Usage

Custom Stock List

# Load your own stock universe
import pandas as pd
stocks_df = pd.read_csv("my_stocks.csv")
stocks = stocks_df["symbol"].tolist()

predictions = predictor.predict(title, stocks, top_k=20)

Batch Prediction

news_headlines = [
    "Nvidia announces new GPU",
    "Tesla deliveries beat estimates",
    "Amazon expands AWS services"
]

all_predictions = []
for headline in news_headlines:
    preds = predictor.predict(headline, all_stocks)
    all_predictions.append({
        "headline": headline,
        "predictions": preds
    })

Custom Candidate Size

# Retrieve more candidates for better recall
predictions = predictor.predict(
    title, 
    all_stocks, 
    top_k=10,
    candidates_k=100  # Default: 50
)

📚 Model Card

Component	Size	Description
Tokenizer	35K vocab	Financial domain tokenizer
Title Encoder	113M	Encodes news headlines
Stock Encoder	113M	Encodes stock symbols
Reranker	113M	Scores title-stock pairs
Total	~340M params	Complete pipeline

🎓 Citation

@misc{stocky2025,
  title={Stocky: Stock Prediction from News Headlines},
  author={Stocky AI Team},
  year={2025},
  publisher={Hugging Face},
  howpublished={\\url{https://huggingface.co/stocky-ai/stocky-stock-predictor}}
}

📄 License

Apache 2.0

🤝 Contact

• Organization: stocky-ai
• Issues: Report bugs or request features on GitHub

---

Built with ❤️ using PyTorch and Transformers