README.md

Semester Project for NI-MVI

This semester project (SP) focuses on "A first exploration of the new minGRU models for time series analysis".

Supervisor: Ippocratis Saltas, ippocratis.s@gmail.com

Detailed Assignment Description

As part of this assignment, implement in PyTorch a minGRU architecture following the specifications from Feng et al. (2024).

We perform a first exploration of the potential of the recently-introduced minGRU models for parameter inference and forecasting in time series. Which consists of three tasks:

1. Parameter inference of sinusoidal waves
2. Time series forecasting on sinusoidal waves
3. Time series forecasting on market stock data

Dataset Information

This project utilizes datasets tailored to the specific requirements of each task. Below is an overview:

1. Parameter Inference (Sinusoidal Waves)

• Description: Synthetic dataset for estimating sinusoidal wave parameters (amplitude, frequency).
• Composition:
  • 5000 samples with 100 data points per sinusoidal wave.
  • Amplitude and frequency randomly varied within predefined ranges. (0-10)

2. Time Series Forecasting (Sinusoidal Waves)

• Description: Synthetic dataset for predicting future points in sinusoidal waves.
• Composition:
  • 5000 samples with 100 input points and 1 target point.
  • Amplitude and frequency randomly varied within predefined ranges. (0-10)

3. Stock Market Forecasting

• Description: Real-world stock data sourced via yfinance.
• Details:
  • IBM stock data (2014-01-01 to 2024-04-01).
  • Daily closing prices as primary feature.
  • Sliding window approach: 60 time steps for input, predicting the next 60.

Requirements

To run this project, you will need the following dependencies:

• Python 3.x
• PyTorch
• torchmetrics
• NumPy
• Matplotlib
• Jupyter Notebook
• Darts
• Prophet

Instructions for Running

• Run the following Jupyter notebook for parameter inference of sinusoidal waves:

  mingru_notebook_param_infer.ipynb

• Run the following Jupyter notebook for time series forecasting of sinusoidal waves:

  mingru_notebook_forecasting.ipynb

This notebook also contains code for loading models from "trained_models" directory.

• Run the following Jupyter notebook for minGRU time series forecasting of stock data:

  stock_data_mingru.ipynb

• Run the following Jupyter notebook for DARTS models time series forecasting of stock data:

  stock_data_darts.ipynb