TimeGrad<T>

Models & Types · AiDotNet.Finance.Forecasting.Foundation

TimeGrad — Autoregressive Denoising Diffusion Model for Time Series Forecasting.

Inheritance: NeuralNetworkBase<T> → FinancialModelBase<T> → ForecastingModelBase<T> → TimeSeriesFoundationModelBase<T> → TimeGrad<T>

For Beginners

TimeGrad predicts time series step by step, where at each step it uses a diffusion process to generate the next value. Think of it as a storyteller who writes one sentence at a time, but for each sentence uses a careful drafting process to get it right. By generating many possible futures, TimeGrad provides not just a single forecast but a range of scenarios with probabilities, helping you understand how confident the prediction is.

How It Works

TimeGrad combines an autoregressive RNN with a conditional diffusion process for probabilistic multi-step forecasting. It generates multiple forecast samples to provide well-calibrated uncertainty estimates.

Reference: Rasul et al., “Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecasting”, ICML 2021.

Example

// Create a TimeGrad autoregressive diffusion model for probabilistic forecasting
// Combines RNN with conditional diffusion for step-by-step forecast generation
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.Regression,
    inputHeight: 512, inputWidth: 1, inputDepth: 1, outputSize: 24);
            
// Training mode with RNN encoder and denoising diffusion decoder
var model = new TimeGrad<double>(architecture);
            
// ONNX inference mode with pre-trained model
var onnxModel = new TimeGrad<double>(architecture, "timegrad.onnx");

Constructors

Constructor Summary
TimeGrad(NeuralNetworkArchitecture<T> architecture, TimeGradOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a TimeGrad model in native mode for training.
TimeGrad(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, TimeGradOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a TimeGrad model using a pretrained ONNX model.

Properties

Property Summary
FoundationModelSize ModelSize { get; }
bool IsChannelIndependent { get; }
bool SupportsTraining { get; }
bool UseNativeMode { get; }
int MaxContextLength { get; }
int MaxPredictionHorizon { get; }
int NumFeatures { get; }
int PatchSize { get; }
int PredictionHorizon { get; }
int SequenceLength { 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)
Tensor<T> ForwardForTraining(Tensor<T> input) epsilon_theta(x_k, h_{t-1}, k) for every future step of a pair from PrepareTrainingPair, as [B, ForecastHorizon]: the RNN reads the teacher-forced sequence and its state before each future step conditions the denoiser.