VKOSPI Multivariate Forecasting Project

This repository contains a complete multivariate time-series forecasting study for the VKOSPI Index, using both classical statistical models and modern deep learning / foundation models. The project is structured into two Jupyter Notebook files, each representing a different modeling paradigm.

📁 Project Structure

project_root/
│── vkospi_multivariate_forecasting.ipynb        # Preprocessing + AR/VAR/ADL + Granger causality
│── vkospi_multivariate_DNN_forecasting.ipynb    # Chronos-2 Zero-Shot + BiLSTM forecasting
│── data/
│     ├── VKOSPI_2004_2025_CL.csv
│     ├── DGS10_2004_2025.csv
│     └── ...
│── module.py
│── train_DL_batch.py
│── README.md

📌 1. vkospi_multivariate_forecasting.ipynb

This notebook contains all classical forecasting workflows, including preprocessing, statistical modeling, and causality analysis.

Included Sections

✓ Data Preprocessing

• Handling missing values
• Log-return construction
• Rate differencing
• Date alignment across all series

✓ AR (Autoregressive) Model

• BIC-based optimal lag selection
• Rolling 1-step-ahead forecasting
• RMSE / MAE / MAPE evaluation

✓ VAR (Vector Autoregression)

• Multivariate interaction modeling
• BIC lag selection
• Rolling-window forecasting for realistic performance evaluation

✓ ADL (Autoregressive Distributed Lag)

• Combines lagged target and lagged exogenous variables
• BIC-based selection for (p_y, p_x)
• Re-fitted at every test step

✓ Granger Causality Analysis

• Pairwise Granger tests
• p-value matrix and significance heatmap

Outputs

• Full forecast curves
• Rolling predictions
• Error metric tables
• Causal relationship visualization

📌 2. vkospi_multivariate_DNN_forecasting.ipynb

This notebook includes deep learning and foundation-model-based forecasting methods.

✓ Chronos-2 Zero-Shot Forecasting

Amazon’s time-series foundation model.

Features:

• No training required
• Probabilistic forecasting (quantiles: 0.1, 0.5, 0.9)
• Iterative 1-step forecasting

Benefits:

• Strong out-of-the-box baseline
• Robust under volatility and regime shifts

✓ BiLSTM Forecasting

A lightweight learned model.

Model:

• Bidirectional LSTM
• Dense output head
• MSE loss with Adam optimizer

Training:

• Sliding-window batching
• Training script: train_DL_batch.py

Outputs:

• Full prediction curves
• Zoom-in comparisons
• RMSE / MAE / MAPE metrics

📊 Model Comparison Summary

Model Category Notes AR Statistical Univariate baseline VAR Statistical Captures multivariate dependencies ADL Statistical Includes lagged exogenous variables Chronos-2 Zero-Shot Foundation Model Strong no-training baseline BiLSTM Deep Learning Learns nonlinear relationships

📈 Evaluation Metrics

All models are evaluated using the following metrics

• RMSE (Root Mean Squared Error)
• MAE (Mean Absolute Error)
• MAPE (Mean Absolute Percentage Error)

🚀 How to Run

1) Install dependencies

pip install -r requirements.txt

2) Place datasets inside data/ directory

The VKOSPI and DGS10 datasets must be available.

3) Open notebooks

Statistical models:

• vkospi_multivariate_forecasting.ipynb

Deep-learning / foundation models: How to Train

python train_DL_batch.py --model_name LSTM --batch_size 128 -- LR 1e-4 --max_epoch 400 --show_train_loss True --use_scaler True --ft_epoch 50 --ft_LR 1e-5 --window_len 300 --d_model 64

Check metrics and plotting

• vkospi_multivariate_DNN_forecasting.ipynb

🧾 Key Findings

•	Chronos-2 provides strong zero-shot performance without training.
•	BiLSTM can match or outperform classical models with sufficient training.
•	VAR and ADL offer interpretability and strong short-term forecasting ability.
•	AR serves as a simple but meaningful baseline.
•	Granger causality reveals important directional effects between financial variables.

🙋 Contact

For questions, feedback, or collaboration inquiries, feel free to open an issue.