Switch-Hurdle: A MoE Encoder with AR Hurdle Decoder for Intermittent Demand Forecasting
Fabian Mușat, Simona Căbuz · 2026
Read Switch-Hurdle: A MoE Encoder with AR Hurdle Decoder for Intermittent Demand Forecasting by Fabian Mușat, Simona Căbuz (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-22685.txt.
Description
Fabian Muşat and one co-author forecast intermittent demand, the retail pattern of long zero-sale stretches broken by sudden orders. Classic tools like ARIMA, exponential smoothing, and Croston variants, plus modern DeepAR and transformer models, all stumble here. They treat demand as one smooth process or grow too costly across many sparse series.
Switch-Hurdle splits the job in two. A mixture-of-experts encoder routes each input to one expert going forward, while staying nearly dense going backward through a straight-through estimator. A hurdle decoder then separates the sale-or-no-sale call from the how-many call, using cross-attention autoregression with a shared hurdle head. On the M5 benchmark and a large retail dataset, it reached leading accuracy while staying scalable.
Catalog record
| CREATOR | Fabian Mușat, Simona Căbuz |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-22685