Parameter-efficient deep probabilistic forecasting

Probabilistic time series forecasting is crucial in many application domains, such as retail, ecommerce, finance, and biology. With the increasing availability of large volumes data, a number neural architectures have been proposed for this problem. In particular, Transformer-based methods achieve state-of-the-art performance on real-world benchmarks. However, these require parameters to be learned, which imposes high memory requirements computational resources training models. To address problem, we introduce novel bidirectional temporal convolutional network that requires an order magnitude fewer than common approach. Our model combines two networks: first encodes future covariates series, whereas second past observations covariates. We jointly estimate output distribution via networks. Experiments four datasets show our method performs par with probabilistic methods, including approach WaveNet, point metrics (sMAPE NRMSE) well set range (quantile loss percentiles) majority cases. also demonstrate significantly means can trained faster lower requirements, consequence reduces infrastructure cost deploying