TimeLLM<T>
Models & Types · AiDotNet.Finance.Forecasting.Foundation
Time-LLM (Large Language Model Reprogramming for Time Series) implementation.
Inheritance: NeuralNetworkBase<T> → FinancialModelBase<T> → ForecastingModelBase<T> → TimeLLM<T>
For Beginners
Time-LLM is a clever way to use powerful language models for time series:
The Key Insight: LLMs like GPT/LLaMA are amazing at pattern recognition in sequences. Time-LLM asks: “Can we make time series ‘speak’ the language of LLMs?”
How It Works:
- Patch Reprogramming: Convert time series patches into “prompt-like” tokens
- Text Prototypes: Learn embeddings that bridge numeric and text domains
- Frozen LLM: The LLM weights stay fixed (no fine-tuning needed)
- Output Projection: Map LLM output back to forecast values
Advantages:
- Leverages powerful pretrained LLMs without expensive fine-tuning
- Works with any LLM backbone (GPT-2, LLaMA, etc.)
- Only trains small reprogramming layers
- Zero-shot transfer to new domains
How It Works
Time-LLM repurposes frozen large language models for time series forecasting by learning a reprogramming layer that translates time series into text-like representations that the LLM can understand.
Reference: Jin et al., “Time-LLM: Time Series Forecasting by Reprogramming Large Language Models”, 2024.
https://arxiv.org/abs/2310.01728
Example
// Create a Time-LLM model that repurposes frozen LLMs for time series forecasting
// Learns a reprogramming layer to translate time series into text-like representations
var architecture = new NeuralNetworkArchitecture<double>(
inputType: InputType.OneDimensional,
taskType: NeuralNetworkTaskType.Regression,
inputHeight: 512, inputWidth: 1, inputDepth: 1, outputSize: 24);
// Training mode with patch reprogramming and frozen LLM backbone
var model = new TimeLLM<double>(architecture);
// ONNX inference mode with pre-trained model
var onnxModel = new TimeLLM<double>(architecture, "time_llm.onnx");
Constructors
| Constructor | Summary |
|---|---|
TimeLLM(NeuralNetworkArchitecture<T> architecture, TimeLLMOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) |
Initializes a new instance in native mode for training. |
TimeLLM(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, TimeLLMOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) |
Initializes a new instance using an ONNX pretrained model. |
Properties
| Property | Summary |
|---|---|
bool IsChannelIndependent { get; } |
|
bool SupportsTraining { get; } |
|
bool UseNativeMode { get; } |
|
int NumFeatures { get; } |
|
int NumPrototypes { get; } |
Gets the number of text prototypes. |
int PatchSize { get; } |
|
int PredictionHorizon { get; } |
|
int SequenceLength { get; } |
|
int Stride { get; } |
|
string LLMBackbone { get; } |
Gets the LLM backbone type. |
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) |