narxpy: PyTorch NARX Implementation

A simple, reusable PyTorch implementation of a Nonlinear Autoregressive Network with Exogenous Inputs (NARX).

Features

• Supports configurable input/output delays (d_i, d_o).
• Handles multi-dimensional exogenous (x) and endogenous (y) time series.
• Configurable hidden layer size and output activation function.
• Supports different simulation modes:
  • Close Loop (Parallel): Uses own predictions for feedback (standard simulation/forecasting).
  • Open Loop (Series-Parallel): Uses provided true values for feedback (teacher forcing, common for training).
• Includes optional bootstrapping for initializing close-loop simulations.

Installation

pip install narxpy

Usage

import torch
from narxpy import NARX

model = NARX(d_i=d_i, d_o=d_o, d_x=d_x, d_y=d_y, d_hl=d_hl, act_func=act_func)

Run Modes

1. Close Loop (Simulation - Default)

y_pred_close = model(x_data, mode="close")
print(f"Close loop output shape: {y_pred_close.shape}")

2. Open Loop (Teacher Forcing)

y_pred_open = model(x_data, mode="open", y=y_true)
print(f"Open loop output shape: {y_pred_open.shape}")

3. Close Loop with Bootstrap

y_pred_bootstrap = model(x_data, mode="close", y=y_true, bootstrap=bootstrap_steps)
print(f"Bootstrap output shape: {y_pred_bootstrap.shape}")