Fracast-0

Fracast-0 is an 85K-parameter pretrained Time Series Foundation Model, small enough to load directly in a web page.

The technical report is available at arXiv:2609.32209. Source code and reproducible pretraining recipes are available in the GitHub repository.

Fracast-0 is a compact pretrained model for probabilistic time-series forecasting. It uses a 2,048-point context, predicts 48 future points, and emits quantiles from 0.1 through 0.9.

The release has 85,001 parameters. The FP32 checkpoint is 344,412 bytes; the W8 checkpoint is 113,228 bytes.

Python Quick Start

Install the Python package in one command:

python -m pip install fracast

The package embeds both checkpoints, so this loads without network access:

import numpy as np
from fracast import FracastModel

model = FracastModel.from_pretrained(device="cpu")

# Input: [time]. Output: [horizon, quantile].
context = np.sin(np.arange(240) / 7.0)
forecast = model.forecast(context)
assert forecast.shape == (48, 9)

# The native head predicts 48 steps; longer horizons roll median blocks.
long_forecast = model.forecast(context, horizon=144)
assert long_forecast.shape == (144, 9)

# Input: [variables, time]. Output: [variables, horizon, quantile].
multivariate = np.stack([context, context * 0.5 + 2.0])
forecast = model.forecast(multivariate)
long_multivariate = model.forecast(multivariate, horizon=144)
assert forecast.shape == (2, 48, 9)
assert long_multivariate.shape == (2, 144, 9)

Pass weights="fp32" for the bundled full-precision checkpoint. The first argument can also be a local directory or a Hugging Face repository id such as "ztxtech/fracast-0" when remote loading is preferred. The W8 file stores symmetric per-output-channel INT8 weights with FP32 scales and residual tensors, then dequantizes them for inference.

Inference Contract

• Input shapes: [T], [1, T], or [V, T].
• Output shapes: [H, 9], [1, H, 9], or [V, H, 9]; H defaults to 48.
• Maximum context: 2,048 observations; longer input keeps the latest 2,048.
• Minimum context: 8 finite observations per channel.
• Missing values: use NaN; the runtime masks them.
• Quantile order: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9].
• Multivariate input is forecast channel-independently. The model does not use cross-variable interactions or future covariates.
• Quantile crossings are not post-processed away.

The browser Playground loads the W8 checkpoint into page memory and releases it when the page is closed.

GIFT-Eval Benchmark

Fracast-0 was evaluated on all 97 GIFT-Eval configurations with the official aggregation protocol. The result is labeled pretrained: the complete pretraining recipe includes datasets from families present in the GIFT-Eval test corpus.

The official aggregate first divides each configuration by the matching Seasonal_Naive result and then takes the geometric mean over all 97 configurations. The leaderboard result uses the FP32 release checkpoint. The shipped W8 form dequantizes weights before inference and is not an integer-only kernel. The official results are public in the GIFT-Eval results/Fracast-0 directory. The compact summary is in benchmark/gift_eval/.

TIME Benchmark

Fracast-0 completed all 98 TIME tasks in a local Apple MPS run with batch size 512 and the FP32 release checkpoint. The normalized scores use seasonal-naive normalization by (dataset_id, horizon) and a geometric mean over MASE and CRPS.

The ranking table is the official 29-model TIME leaderboard, and the raw 98-task outputs are in Real-TSF/TIME-Output. The compact scores in this archive come from the archived raw outputs and the official table; no separate recomputation is recorded for them.

Of the 98 tasks, 47 use the released 48-step head with median-quantile feedback for longer horizons. This rollout path is disclosed in every raw task configuration and is not a strict official-protocol claim. The compact analysis and reproducibility scripts are archived in the GitHub TIME archive and the public tables are also stored here under benchmark/time_benchmark/.

In the parameter-count analysis, Fracast-0 has 85,001 parameters and lies on both the parameter-MASE and parameter-CRPS Pareto fronts. Its dominator lists are empty for both metrics. The smallest better-performing comparator is Toto-2.0-4m with 4,144,456 parameters, MASE 0.691002, and CRPS 0.582490. OmniScient uses a conservative Chronos-2 base lower bound, and PatchTST-FM-Extended uses the public R1 count from the same default architecture.

FEV-Bench

Fracast-0 also completed the official local 100-task FEV-Bench run with Apple MPS, batch size 512, FP32 computation, and the w8 checkpoint. The full adapter, pinned analysis code, raw result, and ranking tables are archived in the GitHub repository. The raw CSV is also stored here as benchmark/fev_bench/fracast-0.csv.

The raw SQL win rate is 44.44% with a 35.72% skill score. All 100 tasks completed without a task failure. Fracast-0 was pretrained on every autogluon/fev_datasets configuration in the benchmark, so the official leakage-controlled aggregate replaces Fracast-0 errors with Chronos-Bolt. This is an in-corpus result and must not be read as an independent zero-shot score. The protocol and comparison table are in the FEV-Bench leaderboard and the FEV-Bench repository.

572 of 235,039 sequence windows (0.243%) had fewer than eight finite observations after official task slicing. The adapter repeats the most recent finite value for those windows; it does not use labels to change forecasts.

BOOM Benchmark

Fracast-0 completed all 7,413 BOOM dataset-term configurations and all 623,602 forecasts with the FP32 checkpoint and no task failures. In the official 24-model comparison, it obtains scaled MASE 0.723, scaled CRPS 0.434, and mean per-dataset rank 11.808. The corresponding ordinal positions are 12 / 24 by MASE and 11 / 24 by CRPS.

Datadog/BOOM is also part of the complete Fracast-0 pretraining recipe, so this result is in-corpus and must not be presented as zero-shot. The official protocol and comparison table are in the BOOM leaderboard. The raw per-configuration result, manifest, and comparator leaderboard are archived in the GitHub repository.

Scenario Highlights

These slices use per-configuration or per-task error ratios against Seasonal Naive. Lower is better, and TIME ranks use the official 29-model table.

• GIFT-Eval, Sales: All four short-horizon Sales configurations improve on Seasonal Naive in normalized MASE and normalized MWQL. The geometric means are 0.693 and 0.422, and Fracast-0 leads TinyCast on both metrics in every configuration.
• GIFT-Eval, Cloud operations: Across the three hourly bizitobs_l2c short, medium, and long slices, normalized MASE is 0.430 and normalized MWQL is 0.353.
• TIME, Solar forecasting: Australia_Solar/H ranks 4th to 6th of 29 by MASE and 7th to 13th by CRPS over short, medium, and long horizons.
• TIME, Manufacturing: On Smart_Manufacturing/H, medium and long horizons rank 8th of 29 by MASE, while CRPS ranks range from 8th to 10th across all three horizons. These are scenario slices, not separate aggregate rankings. The benchmark disclosures above still apply.

Files

Training Data

Fracast-0 was trained on five pinned public datasets plus the four official TinyCast synthetic shards. The five source datasets are immutable snapshots in script/data/download_public_data.py:

The five sources total about 2.6 TiB before conversion. They are converted independently into the fast-corpus layout; the TinyCast synthetic shards are built from the official four-shard recipe and converted into the same layout.

The corpus scan covered 69,614,455 series and about 549.85 billion points across 3,234 datasets; the training split used 62,653,010 series. The checkpoint is from step 45,776 with an effective batch of 4,096 windows, approximately 187.5 million training windows.

Because several corpus roots can overlap GIFT-Eval families, results are labeled pretrained, not zero-shot.

Limitations

• Fracast-0 is univariate and channel-independent; it does not model cross-variable dependence.
• The native head predicts 48 steps at a time. Longer horizons append median predictions and re-normalize each 2,048-point context before the next block.
• The GIFT-Eval result is labeled pretrained with test-data leakage because the pretraining recipe includes GIFT-Eval-family datasets.
• The FEV-Bench pretraining corpus overlaps every benchmark dataset, so its official leakage-controlled result is not an independent zero-shot score.
• 47 of 98 TIME tasks use median-quantile rollout beyond the native 48-step head; the compact scores come from the archived raw outputs and the official 29-model table, with no separate recomputation recorded for them.
• The W8 export is not an integer-only kernel.
• This release is distributed under the Apache-2.0 license.

Acknowledgements

Fracast-0 depends on the teams behind the benchmark, compact baseline, and public pretraining corpora:

• GIFT-Eval provides the 97-configuration evaluation protocol used for the public result.
• FEV-Bench provides the 100-dataset evaluation protocol and leaderboard submission format.
• TIME provides the 98-task evaluation workflow, output format, and leaderboard submission protocol.
• TinyCast provides the compact design baseline and the Apache-2.0 components recorded in the source repository.
• Salesforce/GiftEvalPretrain, Salesforce/lotsa_data, autogluon/chronos_datasets, Datadog/BOOM, and autogluon/fev_datasets provide the public training data.
• The TinyCast team also publishes the synthetic pretraining shards used by the recipe.

Development support

The implementation, debugging, and release work used paid API access to Xiaomi MiMo 2.6 Pro and DeepSeek V4.1 Flash. Both were effective for this project and are recommended.

Supporters / sponsors

There are no sponsors yet. Continued maintenance still needs API credits and server resources. Feedback and support are welcome through:

• zhantianxianguestc@hotmail.com

Citation

@article{zhan2026fracast0fractalweightsharing,
  title={Fracast-0: Fractal Weight Sharing for a Time Series Foundation Model with Only 85K Parameters},
  author={Tianxiang Zhan and Huanyao Zhang and Yuanpeng He},
  year={2026},
  journal={arXiv preprint arXiv:2609.32209},
  eprint={2609.32209},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  doi={10.48550/arXiv.2609.32209},
  url={https://arxiv.org/abs/2609.32209}
}