LoginSign Up
← Back to all sparks
E

echos

INFRA · APIS
Velocity0.0

Echo State Networks for Time Series Modeling and Forecasting

github.com/ahaeusser/echos ↗

Echo state networks for R forecasting, filling in the pieces a fable model is expected to have.

time-seriesforecastingreservoir-computinghyperparameter-tuningfable-ecosystem
◆Current state
echos fits echo state networks, a reservoir-computing approach to time series forecasting, and exposes them through the fabletools model interface so they sit alongside other models in a fable workflow. The three releases in this window take it from a working model to a complete one: forecast intervals in 1.0.2, hyperparameter tuning by rolling-origin cross-validation in 1.0.3, and documentation covering the architecture, hyperparameters and tuning workflow in 1.0.4. Cadence is a few releases a year.
◆Where it's heading
The arc here is a model implementation earning its place in an established framework. Point forecasts came first, then the interval forecasts that any fable-compatible model is expected to produce, generated by bootstrapping residuals and taking quantiles from simulated paths, then the tuning machinery that makes the reservoir hyperparameters usable by people who do not already know what alpha and rho do. Version 1.0.4 spending its whole release on documentation and a clearer dataset name is consistent with that: the remaining barrier is comprehension, not capability.
◆Prediction
With intervals and tuning in place, the natural next step is broader integration with the fable ecosystem, such as handling multiple series or ensembling with other model types. The entries do not indicate whether the maintainer intends to go further into reservoir variants or to stabilise what is here.

◆Recent moves

  1. 2mo ago

    Documentation expanded; M4 dataset renamed

    A documentation release covering the network architecture, key hyperparameters and the tuning workflow, plus a clearer name for the bundled M4 subset with the old name kept working. No functional change, but it addresses the gap that tuning support in 1.0.3 exposed.

    View source ↗
  2. 5mo ago

    tune_esn() tunes reservoir hyperparameters by cross-validation

    Adds rolling-forecast cross-validation over the alpha, rho and the new tau reservoir-scaling parameter, with summary and plot methods for inspecting the results. Reservoir networks are sensitive to exactly these settings, so shipping the tuning loop rather than leaving it to users is what makes the model practical outside a demo.

    View source ↗
  3. 1y ago

    Forecast intervals added via moving block bootstrap

    Brings uncertainty to the forecasts by simulating future paths from a moving block bootstrap of the residuals and taking quantiles from the simulations, alongside a plot method for the reservoir's internal states. This is the release that lets an echo state network behave like any other model in a fable workflow rather than a point-forecast special case.

    View source ↗