Neural Net Time Series

(To be removed) Solve nonlinear time series problem using dynamic neural networks

The Neural Net Time Series app will be removed in a future release. For more information, see Transition Legacy Neural Network Code to dlnetwork Workflows.

For advice on updating your code, see Version History.

Description

The Neural Net Time Series app lets you create, visualize, and train dynamic neural networks to solve three different kinds of nonlinear time series problems.

Using this app, you can:

Create three types of neural networks: NARX networks, NAR networks, and nonlinear input-output networks.
Import data from file, the MATLAB^® workspace, or use one of the example data sets.
Split data into training, validation, and test sets.
Define and train a neural network.
Evaluate network performance using mean squared error and regression analysis.
Analyze results using visualization plots, such as autocorrelation plots or a histogram of errors.
Generate MATLAB scripts to reproduce results and customize the training process.
Generate functions suitable for deployment with MATLAB Compiler™ and MATLAB Coder™ tools, and export to Simulink^® for use with Simulink Coder.

Tip

To interactively build and train deep neural networks for time series tasks, use the Time Series Modeler app. For more information, see Get Started with Time Series Forecasting.

Open the Neural Net Time Series App

Before R2026a: MATLAB Toolstrip: On the Apps tab, under Machine Learning and Deep Learning, click the app icon.
MATLAB command prompt: Enter ntstool.

Algorithms

The Neural Net Time Series app provides built-in training algorithms that you can use to train your neural network.

Training Algorithm Description

Levenberg-Marquardt

Training Algorithm	Description
Levenberg-Marquardt	Update weight and bias values according to Levenberg-Marquardt optimization. Levenberg-Marquardt training is often the fastest training algorithm, although it does require more memory than other techniques. To implement this algorithm, the Neural Net Time Series app uses the `trainlm` function.
Bayesian regularization	Bayesian regularization updates the weight and bias values according to Levenberg-Marquardt optimization. It then minimizes a combination of squared errors and weights, and determines the correct combination so as to produce a network that generalizes well. This algorithm typically takes longer but is good at generalizing to noisy or small data sets. To implement this algorithm, the Neural Net Time Series app uses the `trainbr` function.
Scaled conjugate gradient backpropagation	Scaled conjugate gradient backpropagation updates weight and bias values according to the scaled conjugate gradient method. For large problems, scaled conjugate gradient is recommended as it uses gradient calculations which are more memory efficient than the Jacobian calculations used by Levenberg-Marquardt or Bayesian regularization. To implement this algorithm, the Neural Net Time Series app uses the `trainscg` function.

Update weight and bias values according to Levenberg-Marquardt optimization. Levenberg-Marquardt training is often the fastest training algorithm, although it does require more memory than other techniques.

To implement this algorithm, the Neural Net Time Series app uses the trainlm function.

Bayesian regularization

Bayesian regularization updates the weight and bias values according to Levenberg-Marquardt optimization. It then minimizes a combination of squared errors and weights, and determines the correct combination so as to produce a network that generalizes well. This algorithm typically takes longer but is good at generalizing to noisy or small data sets.

To implement this algorithm, the Neural Net Time Series app uses the trainbr function.

Scaled conjugate gradient backpropagation

Scaled conjugate gradient backpropagation updates weight and bias values according to the scaled conjugate gradient method. For large problems, scaled conjugate gradient is recommended as it uses gradient calculations which are more memory efficient than the Jacobian calculations used by Levenberg-Marquardt or Bayesian regularization.

To implement this algorithm, the Neural Net Time Series app uses the trainscg function.

Version History

collapse all

R2026a: Neural Net Time Series app will be removed

The Neural Net Time Series app will be removed in a future release. Use the Time Series Modeler app instead.

For more information, see Transition Legacy Neural Network Code to dlnetwork Workflows.