DeepAR<T>

Models & Types · AiDotNet.Finance.Forecasting.Neural

DeepAR probabilistic autoregressive forecasting model using LSTM networks.

Inheritance: NeuralNetworkBase<T> → FinancialModelBase<T> → ForecastingModelBase<T> → DeepAR<T>

For Beginners

DeepAR is special because it doesn’t just predict a single value - it predicts a probability distribution. This means you get:

  • A most likely value (the median of simulated futures)
  • A measure of confidence (the spread of those futures)
  • The ability to generate prediction intervals (e.g., 95% confidence bounds)

Key features:

  • Autoregressive: Each prediction depends on previous predictions
  • Probabilistic: Outputs full distributions, not just point forecasts
  • Multi-series: Can learn patterns across many related time series
  • Covariates: Can include additional features like holidays or promotions

How It Works

DeepAR is a probabilistic forecasting model that produces forecast distributions rather than point predictions. It uses autoregressive recurrent neural networks to learn temporal patterns and outputs distribution parameters (e.g., mean and standard deviation for Gaussian).

Reference: Salinas et al., “DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks”, International Journal of Forecasting 2020. https://arxiv.org/abs/1704.04110

The network is the paper’s (DeepARNetwork, shared with DeepARModel and checked against a PyTorch transcription). Each input column is a series; with several, a learned item embedding tells them apart (§3.4). Training maximizes the likelihood of the whole window, input then target (§3.2); forecasts are medians or quantiles of ancestral sample paths.

Example

var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.Regression,
    inputSize: 1, outputSize: 1);   // one series; [batch, LookbackWindow, 1] in, [batch, ForecastHorizon, 1] out
var model = new DeepAR<double>(architecture);
var onnxModel = new DeepAR<double>(architecture, "deepar.onnx");

Constructors

Constructor Summary
DeepAR() Creates a DeepAR with the paper’s defaults for a single series.
DeepAR(NeuralNetworkArchitecture<T> architecture, DeepAROptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a trainable DeepAR.
DeepAR(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, DeepAROptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a DeepAR that runs a pre-trained ONNX model.

Properties

Property Summary
bool IsChannelIndependent { get; } Each input column is its own series: DeepAR is one global model over many series.
bool SupportsTraining { get; }
int PatchSize { get; }
int Stride { get; }

Methods

Method Summary
Dictionary<string, T> Evaluate(Tensor<T> predictions, Tensor<T> actuals)
Dictionary<string, T> GetFinancialMetrics()
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input)
ModelMetadata<T> GetModelMetadata()
Tensor<T> ApplyInstanceNormalization(Tensor<T> input)
Tensor<T> AutoregressiveForecast(Tensor<T> input, int steps)