An Echo State Network library for single and multi-frequency ESN (MFESN) forecasting.

Currently we implement:

• ESN (single reservoir, single frequency):
  • ESN class
  • Ridge CV and fitting, with direct and iterated multistep fit methods
  • Simple forecasting, with direct and iterated multistep forecasting methods
• MFESN (multiple reservoirs, multi-frequency):
  • MFESN class
  • Ridge CV and fitting, with high-frequency (nowcasting) fit methods
  • Simple MF forecasting, with high-frequency nowcasting and forecasting methods
• Reservoir matrix generations utility

IMPORTATN NOTICE: as of 07/2024, LibESN is being consistently refactored to improve peformance, code readibility and functionality. See below for breaking changes!

You can read the full documentation here!
_{(The documentation is automatically built from the latest main commit using pdoc)}

• 09/07/2024 - First commit of v0.2 refactoring. The following are breaking changes:
  • Library name is now libesn, not LibESN, to better conform to Python conventions
  • base_datetime submodule has been changed to datetime
  • base_utils submodule has been split:
    • Data utility functions are now in the datautils submodule
    • ShiftTimeSeriesSplit cross-validation class is now in the validation submodule
  • base_functions submodule has been changed to ufuncs
  • matrix_generator submodule has been changed to matgen
    • matrixGenerator() has been signifcantly changed. In particular, dist doest not accept sparse_ options - the sparsity optional argument automatically handles sparseness of entry-wise distributions (see docs)
  • Documentation pages are officially available, but very early stage

LibESN is based on the Python codebased originally developed for the paper "Reservoir Computing for Macroeconomic Forecasting with Mixed Frequency Data" (UKRI funded project, Ref: ES/V006347/1), available at the following links:

• International Journal of Forecasting (Open Access)

• ArXiv

• ResearchGate