This paper builds on recent research on reframing regression problems in time series prediction as classification problems, discretizing a continuous target space to perform predictions for a fixed set of classes. To address the problem of losing relative distance information between target values (which is a common problem in conventional one-hot encoding), we propose Cumulative Binary Encoding (CBE), which preserves order and magnitude information. To effectively utilize CBE, we propose BinConv, a fully convolutional neural network architecture for probabilistic prediction. We demonstrate that convolutional layers, when combined with CBE, are computationally more efficient and improve prediction performance compared to fully connected layers. Experimental results on standard benchmark datasets demonstrate that BinConv outperforms existing methods in both point and probabilistic prediction, while also providing fewer parameters and faster learning speed.
