RATD Stock Forecasting Pipeline

Project Structure

• Train_TCN/ - TCN model training scripts and results
• Retrival/ - Retrieval scripts and outputs
• Diffusion/ - Diffusion model scripts, configs, and outputs
• dataset/ - Datasets for training and evaluation

The workflow is organized into three main stages:

1. Train TCN Embedding Models

Train Temporal Convolutional Network (TCN) models to generate embeddings for stocks.

Run training:

cd Train_TCN
python train_tcn_all.py         # For all stocks
python train_tcn_industry.py    # For industry-specific
python train_tcn_only.py        # For single stock

• Model weights and embeddings will be saved in Train_TCN/results/.

2. Retrieval Stage

Use the trained TCN models to generate retrieval features for downstream tasks.

Run retrieval:

cd Retrival
python retri_all.py         # For all stocks
python retri_industry.py    # For industry-specific
python retri_only.py        # For single stock

• Retrieval outputs will be saved in the corresponding folders (e.g., Retrival/AMZN_k_n_all/).

3. Diffusion Stage

Feed the retrieval outputs into the diffusion models for final forecasting.

Run diffusion forecasting:

cd Diffusion
python exe_stock_forecasting_all.py         # For all stocks
python exe_stock_forecasting_industry.py    # For industry-specific
python exe_stock_forecasting_only.py        # For single stock