CITRAS-FM
CITRAS-FM (Covariate-Informed Transformer for time series Foundation Modeling) is a tiny (7M-parameter) time series foundation model for covariate-informed zero-shot forecasting. It supports zero-shot forecasting on a single target variable or multiple target variables, with optional support for observed covariates (past-only) and known covariates (available throughout the forecast horizon). It delivers sub-0.1-second CPU inference.
- Code: https://github.com/hitachi-ais/citras-fm
- Paper: CITRAS-FM: Tiny Time Series Foundation Model for Covariate-Informed Zero-Shot Forecasting (EUSIPCO 2026, to appear)
Problem Setting
CITRAS-FM handles three types of input variables:
| Variable | Role | Count | Shape | Available range | Examples |
|---|---|---|---|---|---|
| Target | Variable to forecast | Ct (≥1) | (L, Ct) | Historical window [1, L] | Traffic amount |
| Observed covariate | Exogenous variable influencing the target, available up to the prediction point | Co (≥0) | (L, Co) | Historical window [1, L] | Weather |
| Known covariate | Exogenous variable influencing the target, available from the past through the forecast horizon | Ck (≥0) | (L+H, Ck) | Full window [1, L+H] | Calendar events |
- L: length of historical context
- H: forecast horizon
Covariates are optional — the model works with targets only.
Performance — fev-bench (zero-shot)
fev-bench is a standardized benchmark for zero-shot time series forecasting covering 100 tasks across diverse domains and frequencies. Skill Score is computed relative to Scaled Quantile Loss of the SeasonalNaive baseline (0% = SeasonalNaive, higher is better).
Models with < 10M parameters
| Model | # Params (M) | fev-all (%) ↑ | fev-cov (%) ↑ | fev-multi (%) ↑ | fev-uni (%) ↑ |
|---|---|---|---|---|---|
| CITRAS-FM | 7.2 | 41.2 | 39.0 | 54.2 | 31.3 |
| KAIROS_mini (Ant Group) | 9.9 | 37.7 | 35.4 | 50.9 | 27.9 |
| Chronos-Bolt_Tiny (Amazon) | 8.7 | 35.9 | 32.9 | 49.8 | 26.4 |
| YINGLONG_6m (Alibaba) | 7.3 | 25.1 | 26.5 | 49.7 | -6.2 |
| TinyTimeMixer_r2 (IBM) | 0.8 | -1.1 | -4.6 | 28.2 | -27.5 |
| SeasonalNaive | 0 | 0.0 | 0.0 | 0.0 | 0.0 |
CITRAS-FM achieves the best accuracy among sub-10M tiny TSFMs on all splits, and outperforms several models with 20× more parameters (TimesFM-2.5, COSMIC) in covariate-informed settings. See the GitHub README for the full benchmark tables.
Quick Start
pip install citras-fm
from citras_fm import CitrasFM
model = CitrasFM.from_pretrained("hitachi-nlp/citras-fm")
print(model.pred_len) # 24
print(model.quantiles) # [0.1, 0.2, ..., 0.9]
# Forecast a single 1-D series
import numpy as np
target = np.random.randn(512)
median, quantiles = model.forecast(target, horizon=24)
# median: shape (24,), quantiles: shape (24, 9)
To pin a specific checkpoint revision, pass revision= (an HF commit hash or a git tag):
model = CitrasFM.from_pretrained("hitachi-nlp/citras-fm", revision="ckpt-2026-06-26")
For multivariate / covariate-informed forecasting and fine-tuning, see the example notebooks in the GitHub repository (example_zeroshot.ipynb and example_finetune.ipynb).
Files in this repository
| File | Description |
|---|---|
model.safetensors | Model weights (state_dict, ~29 MB) |
config.json | Hyperparameters consumed by CITRAS.__init__ |
License
- Pretrained weights (this repository): CC BY-NC-SA 4.0.
- Source code (github.com/hitachi-ais/citras-fm): PolyForm Noncommercial License 1.0.0.
Citation
Yamaguchi et al., "CITRAS-FM: Tiny Time Series Foundation Model for Covariate-Informed Zero-Shot Forecasting,"
34th European Signal Processing Conference (EUSIPCO), 2026 (to appear).
Disclaimer
These pretrained weights are provided for research and non-commercial evaluation purposes only, "AS IS", without warranty of any kind, express or implied, including warranties of accuracy, reliability, fitness for a particular purpose, and non-infringement of third-party rights.
In no event shall the authors, contributors, or Hitachi, Ltd. be liable for any damages arising from the use of these weights or any predictions derived from them.
Forecasts produced by this model are probabilistic estimates and may be inaccurate. Do not use this model as the sole basis for safety-critical, mission-critical, life-critical, financial, or medical decisions.