TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity

Analysis

Why This Paper Matters

TinyCast challenges the prevailing trend in time-series forecasting that larger models with attention mechanisms are necessary for zero-shot generalization. By showing that a 146K-parameter model can achieve competitive probabilistic accuracy, it highlights the value of exploiting domain-specific structure—periodicity—rather than relying on massive learned capacity. This is particularly significant for edge deployment, where memory and compute are constrained.

The paper also addresses a gap in the GIFT-Eval benchmark: most zero-shot entries are either large or do not provide predictive distributions. TinyCast is the only sub-1.4M parameter model that emits a distribution without test-data leakage, making it a new reference point for efficient probabilistic forecasting. Its success suggests that for many real-world time series, the periodic structure is more informative than complex learned representations.

Technical Contributions

• Computed Periodicity: A zero-parameter spectral detector identifies dominant periods, which are then used to fold the context on phase. This replaces learned periodicity detection, saving parameters and improving interpretability.
• Attention-Free Architecture: Uses dilated convolutions for encoding and a block-autoregressive quantile decoder, avoiding attention's quadratic complexity and enabling efficient inference.
• Probabilistic Output: Emits a full predictive distribution via quantile regression, not just point forecasts, which is crucial for decision-making under uncertainty.
• INT8 Exportability: The model's operations are limited to convolutions and matrix multiplications, allowing static INT8 quantization and deployment on embedded devices without per-signal fitting.

Results

On GIFT-Eval, TinyCast defines the size-accuracy frontier for probabilistic accuracy among zero-shot entries. It is the only model below 1.4M parameters that emits a predictive distribution and declares no test-data leakage; every entry with better accuracy carries at least that parameter budget. On Chronos-ZS and fev-bench, all neural models ahead of it have at least 28 times its parameters, underscoring its efficiency. While the abstract does not provide exact accuracy numbers, the relative comparisons establish TinyCast's competitive performance at a fraction of the size.

Significance

TinyCast's approach could shift research focus from scaling up models to incorporating domain knowledge and efficient architectures. It demonstrates that careful algorithmic design—like computing periodicity—can outperform brute-force learning for certain tasks. This is especially relevant for IoT and edge AI, where model size and latency are critical. The work also encourages the development of benchmarks that consider parameter efficiency and probabilistic outputs, not just raw accuracy. Future work might explore extending this method to non-periodic series or integrating it with larger models for hybrid approaches.