TimesFM<T>
Models & Types · AiDotNet.Finance.Forecasting.Foundation
TimesFM (Time Series Foundation Model) for zero-shot time series forecasting.
Inheritance: NeuralNetworkBase<T> → FinancialModelBase<T> → ForecastingModelBase<T> → TimeSeriesFoundationModelBase<T> → TimesFM<T>
For Beginners
TimesFM is a revolutionary approach to time series forecasting:
Foundation Model Concept: Just like GPT learns language patterns from vast text data, TimesFM learns time series patterns from millions of diverse series:
- Weather data (temperature, precipitation, wind)
- Financial data (stock prices, exchange rates)
- Retail data (sales, inventory, demand)
- Energy data (consumption, production, prices)
Zero-Shot Capability: The “zero-shot” term means no training required for new tasks:
- Traditional models: Train specifically for your data
- TimesFM: Works immediately on any time series
- Just provide history → get forecasts
Patching Innovation: TimesFM groups consecutive time steps into “patches”:
- Example: 512 time steps → 16 patches of 32 steps each
- Each patch becomes one token for the transformer
- Benefits: Longer context, faster processing, captures local patterns
Decoder-Only Architecture: Like GPT, TimesFM uses causal (one-directional) attention:
- Each position only attends to earlier positions
- Naturally suited for autoregressive forecasting
- Generates predictions step by step
When TimesFM Excels:
- Quick prototyping without model training
- Cross-domain forecasting (same model for different data types)
- Limited historical data (leverages pre-training knowledge)
- Baseline comparisons for specialized models
How It Works
TimesFM is Google’s foundation model for time series forecasting. It uses a decoder-only transformer architecture pre-trained on a massive dataset spanning diverse domains, enabling zero-shot forecasting without task-specific training.
Reference: Das et al., “A decoder-only foundation model for time-series forecasting”, 2024.
https://arxiv.org/abs/2310.10688
Example
// Create a TimesFM (Google) for zero-shot time series forecasting
// Decoder-only transformer pre-trained on diverse domains with patch-based tokenization
var architecture = new NeuralNetworkArchitecture<double>(
inputType: InputType.OneDimensional,
taskType: NeuralNetworkTaskType.Regression,
inputHeight: 512, inputWidth: 1, inputDepth: 1, outputSize: 24);
// Training mode with patch-based decoder-only architecture
var model = new TimesFM<double>(architecture);
// ONNX inference mode with pre-trained model
var onnxModel = new TimesFM<double>(architecture, "timesfm.onnx");
Constructors
| Constructor | Summary |
|---|---|
TimesFM(NeuralNetworkArchitecture<T> architecture, TimesFMOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) |
Creates a TimesFM model in native mode for training or fine-tuning. |
TimesFM(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, TimesFMOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) |
Creates a TimesFM model using 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 NumHeads { get; } |
Gets the number of attention heads in the transformer. |
int NumPatches { get; } |
Gets the number of patches derived from the context length and patch size. |
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) |