FrontierTS - Frontier Time Series Forecasting Model
Architecture
- Type: Decoder-only transformer with patch-based input
- Features: RoPE, SwiGLU, sinh⁻¹ normalization, quantile regression (9 quantiles), multi-token prediction, CPM
- Optimizer: Muon (2x compute efficiency vs AdamW) + AdamW for non-matrix params
Training Data
- autogluon/chronos_datasets (M4, Electricity, Dominick, etc.)
- KernelSynth-style synthetic time series
References
- TiRex (arxiv:2505.23719) | Chronos-2 (arxiv:2510.15821)
- Moirai 2.0 (arxiv:2511.11698) | Moonlight/Muon (arxiv:2502.16982)