GitHub.io Personal Page

Badge Overview

Open Source

Time_Series

This is a repo containing a simple library for Time Series data problems, which tries to aggregate all steps of the analysis, from data preprocessing, to label/target prediction performance assessment, and graphical presentation.

Main Features

• Object oriented API, with a very low learning fixed cost.
• Composable elements, which can be assembled to create a Time Series data pipeline.
• Feature Engineering, like explicit time embeddings (EWMA, Fourier, etc).
• Label transformations for stationarity, such as differencing and Box-Cox transform.
• Outlier detection algorithms (Isolation Forest Detection, Inter-Quartile Range, etc).
• Statistical tests implemented for checking stationarity.
• Both classical and ML time series models wrappers for seamless integration.
• Label/target prediction performance evaluation through Cross Validation.

TODO

• Multiple time steps forecasting for random forests based models.
• Add folder with example notebooks.
• Add Forecast Bias and Normalized Deviation metrics.
• Add TFT model from PyTorch Forecasting
• N-HiTS integration (using PyTorch Forecasting).
• Add automated tests that prove algos are working.
• Add Informer from HuggingFace package. (I don't have enough memory to put the model into GPU memory. Google Colab?)