TimeSeriesTokenizationStrategy
Enums · AiDotNet.Enums
Tokenization strategies for time series foundation models.
For Beginners
Different foundation models use different ways to convert raw time series data into “tokens” that the model can process:
- Patching: Splits the series into fixed-size chunks (most common)
- Quantization: Converts continuous values to discrete vocabulary tokens
- AdaptivePatching: Variable-size patches based on local complexity
- LagFeatures: Uses historical lags as features instead of raw values
- RawSequence: No tokenization — feeds raw values directly
How It Works
Reference: See Chronos (ICML 2024) for quantization, PatchTST (ICLR 2023) for patching, Kairos (2025) for adaptive patching, and Lag-Llama (2023) for lag features.
Fields
| Field | Summary |
|---|---|
TimeSeriesTokenizationStrategy AdaptivePatching |
Variable-size patches based on local information density. |
TimeSeriesTokenizationStrategy LagFeatures |
Historical lag values as input features. |
TimeSeriesTokenizationStrategy Patching |
Non-overlapping or overlapping fixed-size patches. |
TimeSeriesTokenizationStrategy Quantization |
Discrete vocabulary quantization via uniform or learned binning. |
TimeSeriesTokenizationStrategy RawSequence |
Raw numerical sequence without tokenization. |
int value__ |