A time-series foundation model is a large model pretrained on extensive time-series data and reused across forecasting tasks, rather than trained separately for every series. It can simplify forecasting pipelines, lower the entry barrier, and produce useful predictions when little local data is available. section 1.5 section 1.4
Pretraining gives the model broad patterns; fine-tuning or zero-shot inference adapts those patterns to a target series. Robust models must cover varied datasets and frequencies because a model trained at one frequency may fail at another. Generative approaches also support horizons longer than a fixed output window. section 2.9 section 3.9
Common implementations differ in architecture, openness, probabilistic output, and operating requirements:
- TimeGPT provides API-based forecasting, multiple-series prediction, long-horizon forecasting, fine-tuning, and anomaly detection. section 3.0 section 3.9
- Lag-Llama is an open-source probabilistic model using past values and date-time features across daily, hourly, 15-minute, and one-minute frequencies. section 4.6
- Moirai is an open-source probabilistic model with patching, an encoder-only transformer, and native support for exogenous features. section 6.6 section 6.1
Applied evaluation
A practical workflow can forecast weekly sales for four stores using external variables such as holidays, temperature, fuel price, CPI, and unemployment. A capstone evaluation forecasts daily blog visits across 20 cross-validation windows, comparing both accuracy and inference time. section 3.0 section 10.2
These models are powerful baselines, not universal replacements for statistical or data-specific methods. Selection must weigh accuracy, latency, hardware, data availability, privacy, cost, and control. Local models may require GPUs, while proprietary APIs may depend on service availability and paid access. section 10.3 section 3.9 section 1.5
Table 3.6 Pros and cons of foundation forecasting models
| 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 use May not be available if the server is down |
Forecasting on long and short horizons with exogenous features Fine-tuning without local resources |
Compare model strengths, constraints, and suitable forecasting situations before choosing an implementation. section 3.9
