TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity

The Core Claim

Time-series forecasting doesn't need to learn periodicity from scratch. A tiny 146,505-parameter model can outperform much larger zero-shot forecasters by computing dominant periods upfront, then using those periods to restructure the data before feeding it to a compact neural backbone. This approach achieves state-of-the-art size-accuracy tradeoffs and runs on embedded devices with INT8 quantization—no per-signal tuning required.

How It Works

TinyCast splits the forecasting problem into two parts: finding what repeats, and modeling what doesn't.

Finding periodicity: A zero-parameter spectral detector (no learnable weights) analyzes the input signal to identify its dominant periods—the rhythms that naturally recur. This is classical signal processing: look at frequency content, pick the strongest frequencies, convert back to time domain.

Restructuring with phases: Once periods are known, the context window gets "folded" along those phase dimensions. Imagine a daily pattern repeating weekly—instead of processing a flat sequence, you align observations by their phase within each cycle. This makes the periodic structure explicit and easier for downstream layers to exploit.

Neural encoding: A dilated convolutional encoder processes the phase-folded context. Dilated convolutions skip timesteps efficiently, letting the model capture long-range dependencies without ballooning parameters. A block-autoregressive quantile decoder then outputs a full predictive distribution—not just point estimates, but uncertainty bounds at multiple quantiles.

The entire pipeline uses only convolutions and matrix multiplications, no attention mechanisms (which scale poorly). This design choice enables static INT8 quantization: the model compresses to 8-bit integers and runs end-to-end on embedded hardware.

Why It Matters

Parameter efficiency: At 146K parameters, TinyCast is smaller than every zero-shot competitor on GIFT-Eval with published parameter counts. On GIFT-Eval's probabilistic metrics, it defines the frontier—no other model of comparable size emits uncertainty estimates. Among zero-shot models without test-data leakage that output distributions, it's the only sub-1.4M entry; everything scoring better uses at least 1.4M parameters.

Scalable across benchmarks: On Chronos-ZS and fev-bench, every neural baseline ahead of TinyCast carries at least 28× its parameter budget. The gap widens as you move to harder datasets, suggesting the core insight (compute periods, fold context, model residuals) generalizes well.

Deployment viability: Because the model uses only basic operations, it quantizes to INT8 and forecasts entirely on embedded devices—microcontrollers or edge hardware with severe memory constraints. No per-signal fine-tuning, no cloud calls. This matters for IoT, sensor networks, and offline systems where uploading data is infeasible.

Rethinking the learning frontier: The implicit claim is deeper: for periodic signals, computing structure beats learning it when parameter budgets are tight. This challenges the neural-scaling-laws narrative and suggests hybrid approaches (classical + learned) can be more efficient than end-to-end learning for certain domains.