🧠 FinGPT Forecaster

A lightweight Financial Language Model pipeline that leverages GPT-4 to generate training data for forecasting financial asset price movements. This project supports DOW-30, EURO-STOXX-50, and cryptocurrencies, and prepares the data for fine-tuning smaller models like TinyLlama.

---

🚀 Features

• Symbol-wise data preparation with fundamentals (if applicable)
• GPT-4 querying using custom prompts for financial prediction
• Data transformation into LLM-friendly prompt-answer pairs
• Support for multiple indices: DOW-30, EURO-STOXX-50, CRYPTO
• Train/test dataset creation with customizable train ratio
• Ready for fine-tuning models like LLaMA, TinyLLaMA, or other lightweight LLMs

---

📁 Project Structure

1. data/ # Saved raw and processed data
2. prompt.py # Prompt templates for GPT-4
3. data.py # Data preprocessing and GPT querying logic
4. indices.py # List of index symbols
5. main.py # Main launcher script
6. utils/ # (Optional) Utility scripts
7. README.md # This file
8. Fine tuning of Llama using Lora

⚙️ How It Works

1. Data Preparation: Downloads or reads stock/crypto data and optionally includes basic financial metrics.

2. Prompt Generation: Constructs GPT-4 prompts of the form:

3. GPT-4 Querying: Sends prompts to GPT-4 and records its prediction and analysis.

4. Dataset Formatting: Transforms the prompt and answer into instruction-following format with:

• system_prompt
• prompt
• answer

5. Training Ready: Outputs a HuggingFace DatasetDict with train and test splits.

---

🔧 Usage

Run the main pipeline:

python main.py \
--index_name crypto \
--start_date 2022-12-31 \
--end_date 2023-12-31 \
--min_past_weeks 1 \
--max_past_weeks 4 \
--train_ratio 0.6