TimesFM: Time Series Foundation Model for time-series forecasting

Original Article ↗ Hacker News Discussion ↗

A new Google-released “foundation model” for time-series forecasting, TimesFM, is prompting debate over whether generic neural models can meaningfully outperform classic statistical methods like ARIMA, SARIMA, or exponential smoothing. Commenters explore how such a model is pre-trained on large collections of univariate series to learn common patterns (seasonality, trends, momentum), and whether that knowledge usefully transfers to domains like observability, industrial process control, or traffic forecasting. Skeptics question its value for harder domains such as finance, argue that task-specific models and simpler methods often work better and cheaper in practice, and note that similar efforts (e.g., Prophet, Chronos, TimeGPT) have yet to deliver clear real-world advantages.

Scope and Problem Definition

  • Model is focused on univariate, contiguous point forecasts, not complex multivariate setups like machines with thousands of sensors.
  • Some are puzzled by benchmarks and question what real-world series it meaningfully improves on.
  • Unclear whether multiple series submitted together influence each other’s predictions.

How Pretraining on Time Series Might Work

  • Pretraining is described as learning generic temporal patterns: seasonality, trends, momentum, symmetry, and the fact that most series don’t make huge random jumps.
  • Some view this as analogous to pretraining in vision or language: helping models “see” temporal dependencies before task-specific fine-tuning.
  • Interest in whether cross-domain patterns (e.g., physiological signals) can transfer; currently unclear.

Foundation Models vs Task-Specific Models

  • Strong skepticism about “time series foundation models” as a general class, compared to domain-specific or task-specific models.
  • Others argue that multi-task learning and the ubiquity of structured temporal patterns justify exploration.
  • Debate on whether good general forecasting requires an implicit “world model”; some think next-step prediction can induce this, others strongly disagree.

Comparison to Classical and Other ML Approaches

  • TimesFM is compared to Prophet, ARIMA, SARIMA, exponential smoothing, and hybrid methods.
  • Several comments claim classical or hybrid statistical methods often match or beat ML in accuracy and cost for practical forecasting.
  • Deep-learning forecasting models (including Chronos, TimeGPT) are criticized as slower and not clearly better in practice, though some note that this is what was once said about early vision models.

Use Cases and Applications

  • Proposed uses: mouse movement prediction to reduce perceived lag, gameplay and cheat detection, industrial process control, traffic and demand forecasting, observability/anomaly detection, GPS track smoothing.
  • For industrial time series, some report ML not outperforming statistical process control; others point to successful task-specific ML in complex systems (e.g., reactors).

Finance and Stocks

  • Multiple comments doubt usefulness for stock prediction, referencing randomness and efficient-market arguments.
  • Consensus that if a model reliably beat markets, it would not be publicly released.

Tooling and Practical Concerns

  • Requests for a simple chat-style interface to upload series and experiment.
  • Questions about latency, data requirements, and how many metrics per core can realistically be supported are raised but largely unanswered.