Introduction

In MATLAB, neural-network time-series prediction usually means using past values of a sequence to predict future values. The practical challenge is not only choosing a network, but also structuring the time series into the delayed input format the network expects and deciding whether you want one-step-ahead prediction or closed-loop multi-step forecasting.

The Typical MATLAB Approach

A common classic workflow uses timedelaynet from MATLAB's neural network tooling. The network takes previous timesteps as input and predicts the next value.

A minimal example:

1series = sin(0:0.1:20);
2inputSeries = num2cell(series);
3
4delays = 1:2;
5hiddenSize = 10;
6net = timedelaynet(delays, hiddenSize);

This says:

• use the previous 1 and 2 timesteps as input context
• learn a hidden representation with 10 neurons

The choice of delays is important because it determines how much history the model can use.

Preparing the Data

MATLAB neural-network functions usually expect cell-array sequence formats rather than plain numeric vectors. The helper preparets aligns inputs, targets, and internal delay states correctly.

[Xs, Xi, Ai, Ts] = preparets(net, inputSeries, {}, inputSeries);

Then train:

net = train(net, Xs, Ts, Xi, Ai);

And predict:

y = net(Xs, Xi, Ai);

The preparets step is what often confuses new users. Without it, the delayed network inputs and targets do not line up correctly.

One-Step Prediction vs Multi-Step Forecasting

There are two related but different tasks:

• one-step prediction, where the model predicts the next value using true historical data
• multi-step forecasting, where predictions feed back into future predictions

One-step prediction is easier and often looks better in evaluation because the model always sees real history. But for actual forecasting, you usually care about what happens when the model must rely on its own previous outputs.

In MATLAB, that is where closed-loop conversion becomes useful.

Closed-Loop Forecasting

After training an open-loop delay network, you can convert it for recursive forecasting:

netc = closeloop(net);
[Xsc, Xic, Aic, Tsc] = preparets(netc, inputSeries, {}, inputSeries);
yc = netc(Xsc, Xic, Aic);

In closed loop, the network's own outputs are fed back as future inputs. This is more realistic for long-horizon forecasting but also more error-prone because mistakes accumulate over time.

That difference is critical. A model that performs well on one-step prediction can degrade quickly in closed-loop mode.

Choosing Delays and Network Size

There is no universal best setting, but a sensible process is:

1. start with short delays such as 1:2 or 1:5
2. choose a modest hidden layer size
3. evaluate validation error
4. increase complexity only if the simpler model underfits

Using too many delays or too many hidden units can overfit small time series. Neural networks are flexible, but that flexibility can become a liability when the dataset is limited.

Data Splitting and Evaluation

For time series, keep the temporal order intact. Do not randomly shuffle samples before splitting into training and test sets. A realistic evaluation trains on earlier data and validates on later data.

Also compare the neural network against simple baselines such as:

• previous-value persistence
• moving average
• linear autoregression

If the neural network cannot beat a simple baseline, the issue may be feature design, insufficient data, or an unnecessarily complex model.

When Neural Networks Help

Neural networks are most useful when the series contains nonlinear structure, regime changes, or interactions with other input variables. For simple smooth series, classical methods can be easier to tune and explain.

So the real question is not “can MATLAB do time-series prediction with a neural network?” It can. The better question is “does the data justify using a neural network instead of a simpler forecasting model?”

Common Pitfalls

• Feeding raw numeric arrays directly into MATLAB neural-network functions without converting them to the sequence format they expect.
• Evaluating only one-step-ahead predictions and assuming the model will perform equally well in closed-loop forecasting.
• Choosing large delay windows and hidden layers immediately, which often overfits limited time-series data.
• Randomly shuffling time-series samples before validation, which leaks future information into training.
• Skipping simple baselines and therefore not knowing whether the neural network is actually adding value.

Summary

• MATLAB supports neural-network time-series prediction effectively with tools such as timedelaynet.
• 'preparets is the standard way to align delayed inputs and targets for training.'
• One-step prediction and closed-loop forecasting are different tasks and should be evaluated separately.
• Delay selection, hidden layer size, and time-aware validation matter more than raw model complexity.
• Always compare against simple forecasting baselines before assuming the neural network is the right model.