The short answer: neither extreme. In 2026, most production forecasting stacks use multiple model families, choosing based on the data regime rather than replacing everything with transformers.
Here's how the landscape has settled.
| Situation | Common choice |
|---|
| Small datasets (hundreds to a few thousand observations) | ARIMA, ETS, Prophet, gradient boosting with lag features |
| Medium tabular forecasting | LightGBM/XGBoost + engineered lags, N-BEATS, N-HiTS |
| Large collections of related time series | Temporal Fusion Transformer (TFT), PatchTST, TiDE, TimeMixer, TimesFM-style foundation models |
| Zero-shot forecasting | Pretrained foundation models (TimesFM, Chronos, Moirai, Lag-Llama, etc.) |
| Mission-critical forecasting | Ensembles of statistical + deep models |
Prophet
Still used, but much less than around 2018–2021.
It's valuable when:
- business users need interpretable trend and seasonality
- holidays matter
- datasets are relatively small
- rapid deployment is more important than squeezing out the last few percent of accuracy
Its limitations become obvious when:
- multiple interacting seasonalities
- many correlated series
- nonlinear relationships
- exogenous variables
- long-horizon forecasting
ARIMA
Definitely not dead.
Many companies still use ARIMA/SARIMA because:
- it's robust
- fast to train
- explainable
- often surprisingly competitive on stable economic or industrial signals
A common benchmark is still:
"If your fancy deep model can't beat SARIMA, don't ship it."
N-BEATS / N-HiTS
These remain among the strongest classical deep-learning models.
Advantages:
- works well on univariate forecasting
- doesn't require attention mechanisms
- relatively data efficient
- often beats transformer models on medium-sized datasets
N-HiTS especially remains a very competitive baseline.
Transformers
Transformers have become mainstream, but not every transformer.
Common architectures include:
- PatchTST
- iTransformer
- TFT (Temporal Fusion Transformer)
- Informer (less popular now)
- Autoformer (less common than before)
- FEDformer
- TimeMixer
- TiDE
The biggest trend is patch-based or efficient transformer architectures rather than vanilla attention over every timestamp.
Foundation models
This is probably the biggest change since 2023.
Models like:
- Google's TimesFM
- Amazon's Chronos
- Salesforce's Moirai
- Lag-Llama
are pretrained on millions or billions of time points.
Benefits:
- zero-shot forecasting
- few-shot fine-tuning
- strong performance when historical data is limited
- easy deployment across thousands of related series
For many organizations these have become the first baseline instead of Prophet.
What companies actually use
Typical production pipeline:
- Naive seasonal baseline
- ETS / SARIMA
- LightGBM with lag features
- N-HiTS or PatchTST
- Foundation model
- Ensemble of the best performers
The ensemble often wins.
If I were building a new forecasting system today
I'd probably evaluate models in roughly this order:
- Seasonal naive baseline
- LightGBM/XGBoost with lag and calendar features
- N-HiTS
- PatchTST
- TimesFM or Chronos as a pretrained baseline
- Ensemble the top performers
I'd only choose Prophet as the primary model if interpretability, holiday effects, and quick deployment outweighed the need for state-of-the-art accuracy.
Overall trend: the field has shifted from "Prophet vs. ARIMA" to "foundation models + specialized neural architectures + strong statistical baselines." Statistical models remain indispensable baselines, while transformers and pretrained time-series models dominate when you have many related series or need scalable, high-accuracy forecasting.
