MOMENT<T>

Models & Types · AiDotNet.Finance.Forecasting.Foundation

MOMENT (Multi-task Optimization through Masked Encoding for Time series) foundation model.

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

For Beginners

MOMENT is the first true multi-task time series foundation model:

Architecture Overview:

  1. Patch Embedding: Divides the input into fixed-length patches (e.g., 64 steps each)
  2. RevIN: Reversible Instance Normalization handles different scales automatically
  3. T5 Encoder: A stack of transformer encoder layers processes the patches
  4. Task Heads: Separate output heads for each supported task

Multi-Task Capability:

  • Forecasting: Linear projection from encoder output to future values
  • Anomaly Detection: Reconstruction-based — anomalies have high reconstruction error
  • Classification: Pooled encoder output fed to a classification head
  • Imputation: Masked patches reconstructed using unmasked context
  • Embedding: Mean-pooled encoder output serves as the representation

Key Insight: MOMENT’s pretraining uses masked reconstruction (like BERT), which naturally supports all five tasks without task-specific pretraining.

Model Sizes (MOMENT family):

  • MOMENT-Small: ~40M parameters
  • MOMENT-Base: ~385M parameters (default)
  • MOMENT-Large: ~1B+ parameters

How It Works

MOMENT is a family of open time series foundation models from Carnegie Mellon University. It uses a T5-based encoder-only transformer with patch-based input and RevIN to handle five downstream tasks: forecasting, anomaly detection, classification, imputation, and embedding generation — all from a single pretrained backbone.

Reference: Goswami et al., “MOMENT: A Family of Open Time-Series Foundation Models”, ICML 2024. https://arxiv.org/abs/2402.03885

Thread Safety: This class is NOT thread-safe. Create separate instances for concurrent usage.

Example

// Create a MOMENT multi-task time series foundation model
// Supports forecasting, anomaly detection, classification, imputation, and embedding
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.Regression,
    inputHeight: 512, inputWidth: 1, inputDepth: 1, outputSize: 24);
            
// Training mode with T5-based encoder and task-specific heads
var model = new MOMENT<double>(architecture);
            
// ONNX inference mode with pre-trained model
var onnxModel = new MOMENT<double>(architecture, "moment_base.onnx");

Constructors

Constructor Summary
MOMENT(NeuralNetworkArchitecture<T> architecture, MOMENTOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a MOMENT model in native mode for training or fine-tuning.
MOMENT(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, MOMENTOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a MOMENT model using a pretrained ONNX model.

Properties

Property Summary
FoundationModelSize ModelSize { get; }
IReadOnlyList<TimeSeriesFoundationModelTask> SupportedTasks { get; }
TimeSeriesFoundationModelTask CurrentTask { 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) Captures per-layer activations of the native forward pass.
ModelMetadata<T> GetModelMetadata()
Tensor<T> ApplyInstanceNormalization(Tensor<T> input)
Tensor<T> AutoregressiveForecast(Tensor<T> input, int steps)
Tensor<T> Classify(Tensor<T> series, int numClasses)
Tensor<T> DetectAnomalies(Tensor<T> series, double? threshold)
Tensor<T> Embed(Tensor<T> series)
Tensor<T> Impute(Tensor<T> series, Tensor<T> mask)