Time Series Foundation Models: Learning Patterns or Universal Temporal Laws?

1. From Single-Sequence Forecasting to Cross-Sequence Learning

Traditional time series forecasting builds a dedicated model for each individual series: one time series → train model → predict future. Foundation models invert this workflow: multiple time series → pre-train a single model → input new series → predict future. This shift relies on transfer learning : a model pre-trained on thousands to millions of diverse sequences can recognize familiar temporal regularities in a never-before-seen series and produce forecasts without any task-specific retraining — a capability called zero-shot forecasting .

Models such as TimesFM , Chronos , Moirai , and timestral have already demonstrated the practical effectiveness of this paradigm.

Multiple time series → Pre-train one model → Input new series → Predict future

When confronted with a new sequence, the model may already be familiar with components such as:

Trends

Daily, weekly, or yearly seasonality

Short-term autocorrelation

Changing variance

Periodic or other repeating patterns

Regime changes (state transitions)

2. What Do Foundation Models Actually Learn?

During large-scale pre-training, the model learns to map historical observations through a temporal representation to future predictions. This representation likely captures several transferable temporal behavior patterns:

Local dynamics: Given a specific pattern, what typically follows?

Seasonality: Does a behavior repeat on daily, weekly, or yearly cycles?

Long-range dependencies: Do earlier events still influence future predictions?

State transitions: Has the sequence switched to a different behavioral mode?

Inter-variable relationships: When one series changes, how do others tend to respond?

These regularities can span application domains. For example, electricity demand and website traffic have completely different business meanings, yet both exhibit cyclic variations, trends, sudden spikes, and regime shifts. Therefore, the model may not learn what a particular series represents in a specific business context, but rather the more general law: how temporal patterns typically evolve over time .

3. Why Does Large-Scale Cross-Sequence Pre-Training Work?

Different time series are generated by distinct latent processes, yet these processes are not entirely unrelated. They may originate from a broader population of temporal behaviors and consequently share certain structural commonalities. Traditional models primarily learn the dynamics of a single dataset. Foundation models, by exposure to many generative processes, gain the opportunity to induce their shared regularities. This is closely related to meta-learning : the model not only learns how to predict a given dataset but also learns the distribution of forecasting problems so that experience can be transferred to new tasks.

4. From Forecasting Toward Understanding System Evolution

If a model masters transferable laws of how systems change over time, forecasting may be just one application. The primary question today is "What will happen next?" However, models with richer temporal representations could also address:

Is the current behavior anomalous? Has the system entered a new operating regime? How would the outcome change under a different scenario? If we actively intervene on a factor, how will the system respond?

At present, the main challenge remains accurately predicting how a system typically operates, which is not equivalent to understanding the consequences of active interventions . Future foundation models may need to progress from recognizing temporal patterns to mastering the underlying dynamics that generate those patterns . Only by crossing that boundary could they take on tasks beyond forecasting. One plausible direction is evolving into general dynamical system models .

Summary

Time series foundation models change how forecasting models are trained and used: learn universal temporal structures from massive datasets, then transfer that knowledge to new forecasting tasks. TimesFM, Chronos, Moirai, and timestral show that this approach already delivers strong zero-shot and few-shot performance across diverse dataset types.

Author: Francisco Mandujano-Reyes