4-Sight

A Multimodal Commodity Price Prediction & Pattern Recognition Tool

4-Sight is a research tool for visualizing and predicting commodity price trends (stocks, ETFs, etc.) through a multimodal approach. It integrates historical OHLCV data with technical chart analysis to detect candlestick patterns and forecast trends.

The system combines:

• Numerical Time-Series Models – leveraging LSTMs and statistical methods (SMA, EMA) on OHLCV data.
• Visual Chart Models – using a Vision Transformer (ViT) trained on annotated candlestick charts with support/resistance and breakout levels.

---

📂 Project Structure & Key Files

1. alphavantage.v1.py

Purpose: Fetch and preprocess raw stock market data.

Key Features:

• Integrates with Alpha Vantage API to download historical OHLCV data.
• Saves data to CSV for later processing.
• Outputs a clean pandas DataFrame ready for downstream tasks (chart generation, LSTM training).

---

2. Pattern Identifier Annotated.py

Purpose: Detect candlestick patterns and generate annotated charts.

Key Features:

• Uses TA-Lib to identify candlestick patterns (with a default window size of 20).
• Generates candlestick charts with mplfinance and matplotlib.
• Highlights detected pattern candles with bounding boxes.
• Draws a trendline for the next 5 candles after the detected pattern.
• Saves charts as .png images, with filenames named after the detected pattern (e.g., CDL3INSIDE_bearish_205.png).
• Exports image_labels.csv mapping filenames to pattern labels for use in supervised learning.

---

📊 Example Output

Example chart with pattern annotation and trendline:

• The black rectangle marks the detected candlestick pattern.
• The dashed blue line indicates the price trend for the following 5 candles.
• File is saved with the pattern label in the name for dataset preparation.

Inference:

---

🔑 Current Focus

• Building a pattern-labeled chart dataset automatically using TA-Lib + matplotlib.
• Each chart = one detected candlestick pattern + future trendline.
• These labeled charts form the vision branch for training the Vision Transformer.

---

🚀 Next Steps

• Train a Vision Transformer (ViT) on the generated chart dataset.
• Fuse predictions from ViT with an LSTM model trained on historical OHLCV values.
• Extend ensemble with SMA/EMA baselines for comparison.