Time Series Forecasting Benchmark

This project explores the benchmarking of univariate time series forecasting, focusing on comparing the performances of various models. The models analyzed include traditional statistical models such as ARIMA and SARIMA, as well as modern deep learning models like LSTM and Transformer networks. Additionally, the study delves into a new class of emerging technologies, continuous deep learning models, with a particular emphasis on Liquid Time Constant Networks and their Closed-form Approximation.

The benchmark is conducted using a simple univariate time series dataset, serving as a reference for all models. Beyond evaluating the inference performance of each model, the analysis also considers the advantages and trade-offs of using one class of models over others for similar forecasting tasks. This comprehensive approach provides insights into the conditions under which certain models may outperform others, facilitating more informed decision-making in model selection for time series forecasting.

---

Requirements

Component	Version
Python	3.12.5
pip	22.0.4

Set Up the Python Environment

python -m venv data-analytics

Download the Dependencies

python -m pip install -r requirements.txt

---

Download the Datasets and Run the notebook

Clone the GitHub repo and change into the repo's directory:

git clone https://github.com/AndrewDarnall/TSF-Benchmark.git
cd TSF-Benchmark

Create the datasets directory:

mkdir datasets
cd datasets

The following notebook uses the AirPassengers dataset, which once downloaded, extract the .zip file into the datasets/ directory

mv ~/Downloads/archive.zip .
unzip archive.zip
rm archive.zip
cd ..

From the commandline terminal (in Linux and macOS) run the following command:

jupyter-notebook ./TSF-Benchmark.ipynb --port=8086