mijanr / forecastflow

An automated code-base for time-series forecasting

ForecastFlow: An automated code-base for time-series forecasting

This repository contains the code for time-series forecasting using various models, such as InceptionTime, LSTM, XGBoost, etc. It contains the complete pipeline from data preprocessing, model building, training, hyperparameter tuning, and evaluation.

MLFlow tracks the experiments and logs the metrics, parameters, and artifacts. Hydra manages configurations, and Optuna optimizes hyperparameters.

The code is designed to be modular and scalable. You can easily add new models, preprocessors, and data loaders!

Data

The data used in this project is the Beijing Multi-Site Air Quality from the UCI Machine Learning Repository. It contains hourly air quality data from 12 different sites in Beijing. The data is preprocessed and aggregated to hourly/daily/weekly data for forecasting purposes.

Models

The following models are used for forecasting:

• XGBoost
• LSTM
• InceptionTime
• etc.

Results

The following are the results of the forecasting models:

• InceptionTime with hourly data, past_horizon = 10, forecast_horizon = 1:

• LSTM with hourly data, past_horizon = 10, forecast_horizon = 1, with Standardized data:

• XGBoost with hourly data, past_horizon = 10, forecast_horizon = 1:

Different windows were used for forecasting purposes in the case of XGBoost, so the plots are not comparable.

Requirements

requirements.yml contains the required libraries to run the code. You can install them using the following command:

conda env create -f requirements.yml

To update with new libraries:

conda env export --no-builds | grep -v "prefix" > requirements.yml