NLARNN

NLARNN, short for nonlinear autoregressive neural network, is a type of neural network used for modeling and forecasting nonlinear time series. It is a member of the autoregressive neural network family and is often implemented as a feedforward network that receives a vector of past observations as input and produces a forecast of the next value.

Typically, the input to a NLARNN consists of the previous p values y(t-1), ..., y(t-p). In some formulations, exogenous inputs u(t-1), ..., u(t-q) may be included, generalizing to a nonlinear autoregressive with exogenous inputs (NARX) structure. The network learns a mapping from the delayed inputs to the current output, capturing nonlinear dependencies in the data.

During training, supervised learning is used with a set of input-output pairs constructed from historical data. Backpropagation or backpropagation through time can be employed, depending on whether the network preserves temporal dynamics through delays or is treated as a static mapping. Model selection involves choosing the number of delays, the size of the hidden layer, activation functions, and regularization to prevent overfitting.

NLARNN is used in various domains requiring nonlinear time series forecasting, including finance, engineering, environmental modeling, and control systems. It can be advantageous when linear models fail to capture nonlinear patterns but can be sensitive to noise and data length.

Relation to other models: NLARNN is closely related to NARX and to nonlinear autoregressive models; when exogenous inputs are omitted, it reduces to a nonlinear autoregressive network. Modern recurrent neural networks and sequence-to-sequence models offer alternative approaches for time-dependent tasks.