TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity
TinyCast claims probabilistic zero-shot forecasting from just 146,505 parameters by computing periodicity instead of learning it.
The model uses a zero-parameter spectral detector to find dominant periods, then folds context by phase before convolutional encoding and quantile decoding. The paper says it is smaller than every countable zero-shot entry on GIFT-Eval and sets the size-accuracy frontier for probabilistic accuracy. It also reports that stronger neural models on Chronos-ZS and fev-bench use at least 28 times more parameters. Its convolution-and-matrix path lets it export to static INT8 and run end to end on embedded hardware without per-signal fitting. HF Daily Papers' note
The model uses a zero-parameter spectral detector to find dominant periods, then folds context by phase before convolutional encoding and quantile decoding. The paper says it is smaller than every countable zero-shot entry on GIFT-Eval and sets the size-accuracy frontier for probabilistic accuracy. It also reports that stronger neural models on Chronos-ZS and fev-bench use at least 28 times more parameters. Its convolution-and-matrix path lets it export to static INT8 and run end to end on embedded hardware without per-signal fitting. HF Daily Papers' note
score 5