Time series and forecasting · insight-0243

Make a foundation or deep forecaster beat seasonal naive and ETS on your backtest first

Situation
A team plans to replace its forecasting stack with a transformer-based model after a promising demo.
Action
Run the candidate, seasonal naive, ETS or ARIMA, and a tuned gradient-boosting model through the same origins and horizons; adopt the complex model only if the gain is material, consistent across segments and worth its cost.
Reason
Simple linear models outperformed several transformer forecasting models on standard long-horizon benchmarks, showing that novelty is not evidence of suitability.
Boundary
Any proposal to adopt deep or foundation forecasting models.

Symptom or risk

Complex models can lose to simple baselines on many series, while adding GPU cost, latency and opaque failure modes.

Large panels with rich covariates or a need for zero-shot cold starts may justify them even at similar accuracy.

Zero-shot models save training effort but limit control over covariates and retraining.

Comparing against an untuned or misconfigured baseline.

Simple linear models (DLinear, NLinear) outperformed several transformer-based long-term forecasting models on standard benchmarks.
Are Transformers Effective for Time Series Forecasting?