Time-series forecasting predicts future values of one or more variables indexed over time. It supports applications such as weather forecasting, demand planning, store-sales prediction, and website-traffic planning. section 1.0 section 10.3
Foundation forecasting models extend this task beyond data-specific models: one model can forecast series with different frequencies, seasonality, holiday effects, and temporal properties. They can also support anomaly detection and time-series classification. section 1.0
Common implementation routes include:
- Use specialized models such as TimeGPT, Lag-Llama, Chronos, Moirai, or TimesFM for zero-shot or adapted forecasting. section 1.4 section 10.3
- Frame forecasting as a language task by formatting an input series as a prompt and extracting predicted values from the model’s output. section 8.0
- Reprogram an LLM for forecasting by aligning time-series patches with text prototypes, accepting that training is required. section 9.6
Prompt-based forecasting
A prompt-based workflow converts numerical observations into natural language input, asks an LLM for the next values, and parses its string output into predictions. Few-shot and chain-of-thought prompts may improve results, but preprocessing and extraction add complexity. section 8.0 section 8.10
Table 8.3 Pros and cons of each foundation forecasting model
| Model |
Pros |
Cons |
When to use |
|---|---|---|---|
| TimeGPT |
Easy and fast to use Comes with many native functions Works on any device, regardless of hardware Free plan |
Paid product for a certain usage Model may not be available if the server is down. |
Forecasting on long and short horizons with exogenous features. You need to fine-tune but do not have the local resources. |
| Lag-Llama |
Open source model Free to use |
Awkward to use because we must clone the repository Speed of inference depends on our hardware. |
Quick proof of concept Ideal for research-oriented projects
|
| Chronos |
Open source model Free to use Can be installed as a Python package |
Speed and accuracy depend on our hardware. Fine-tuning requires cloning the repository. We must define some functions manually. |
Forecasting on horizons shorter than 64 time steps Forecasting series without strong trends |
| TimesFM |
Open source model Free to use Can be installed as a Python package Supports exogenous features |
Only the largest model is available. Restrictive requirements for use (16 GB of RAM, Linux-based system to install dependencies) Knowledge of JAX is required for fine-tuning. Cannot perform anomaly detection because it is a deterministic model |
Ideal for deterministic forecasting in which prediction intervals are not required You want to feed a long input series. |
| Moirai |
Open source model Free to use Supports exogenous features Can be installed as a Python package |
Speed and accuracy depend on the hardware. We must define helper functions when not working entirely with GluonTS. |
You have exogenous features. Forecasting on horizons shorter than 256 time steps |
| Large language models (Flan-T5 and Llama-3.2) |
Possible to use free LLMs Interacting through natural language feels intuitive, |
Not built to handle time-series data Requires lengthy code to preprocess the input and extract the values from the output Requires a lot of computation power for local setups |
You already have access to an LLM. You need a natural language interface for forecasting. You know how to develop prompts and safeguards for edge cases. |
Compare forecasting models by ease of use, hardware requirements, openness, speed, horizon, fine-tuning needs, and intended use. section 8.9
