TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity

Hugging Face Daily Papers Papers
zero-shot time-series foundation-model probabilistic-forecasting attention-free edge-ai small-model

Summary

TinyCast is a compact zero-shot time series foundation model with computed periodicity, enabling efficient probabilistic forecasting on edge devices like Cortex-M7.

We introduce TinyCast, an attention-free zero-shot forecaster that emits a predictive distribution from 146,505 parameters, on the premise that at this size the periodic structure of a context is worth computing rather than learning. A zero-parameter spectral detector supplies the dominant periods, the context is folded on their phase, and a dilated convolutional encoder and a block-autoregressive quantile decoder model the rest. It is smaller than every zero-shot entry on the GIFT-Eval board whose parameter count can be established. On probabilistic accuracy it defines the size-accuracy frontier. Among zero-shot entries declaring no test-data leakage it is the only one below 1.4M parameters that emits a predictive distribution, and every entry scoring better carries at least that budget. On Chronos-ZS and fev-bench every neural model ahead of it carries at least 28 times its parameters. Because the mixing path is convolutions and matrix multiplications only, it exports to static INT8 and forecasts end to end on an embedded device without per-signal fitting.
Original Article
View Cached Full Text

Cached at: 08/21/26, 08:11 PM

Paper page - TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity

Source: https://huggingface.co/papers/2608.15767 TinyCast is a 146,505-parameter zero-shot time series foundation model. It isthe smallestmodel on the GIFT-Eval board with a public per-configuration result and no declared test-data leakage, and below 1.4M parameters it is the only zero-shot entry that emits apredictive distributionrather than point forecasts.

It isattention-free: dilated causal convolutions plus a zero-parameter spectral detector that computes each context’s periodicity instead of learning it, so no capacity is spent rediscovering seasonality.

Because every learned operation is a convolution, a matrix multiplication or a normalization, it exports to static INT8 andruns a full forecast end to end on a Cortex-M7in 4.08 s within 731 KB of RAM, at a cost of about 2% of point accuracy.

Weights, code and the training recipe are public.

Similar Articles

Unified Zero-Shot Time Series Forecasting: A Darts Foundation

arXiv cs.LG

Darts, a popular open-source Python library for time series analysis, introduces a unified FoundationModel class collection that integrates multiple time series foundation models (Chronos-2, TimesFM 2.5, TiRex, PatchTST-FM) for zero-shot and fine-tuned forecasting with standardized interfaces and minimal dependencies.

A decoder-only foundation model for time-series forecasting

Papers with Code Trending

This article presents a research paper on Time-Series Foundation Model (TimeFM), a decoder-only model that achieves near-optimal zero-shot performance across diverse time-series datasets by adapting large language model techniques.