SuperDeConFuse: A supervised deep convolutional transform based fusion framework for financial trading systems

This work proposes a supervised multi-channel time-series learning framework for financial stock trading. Although many deep models have recently been proposed in this domain, most of them treat the trading data as 2-D image data, whereas its true nature is 1-D data. Since systems are existing techniques treating not suggestive any technique to effectively fusion information carried by multiple channels. To contribute towards both these shortcomings, we propose an end-to-end inspired previously established (unsupervised) convolution transform framework. Our approach consists processing channels through separate layers, then fusing outputs with series fully-connected and finally applying softmax classification layer. The peculiarity our framework, that call SuperDeConFuse (SDCF), remove nonlinear activation located between layers well one latter output We compensate removal introducing suitable regularization on aforementioned layer filters during training phase. Specifically, apply logarithm determinant break symmetry force diversity learnt transforms, enforce non-negativity constraint mitigate issue dead neurons. results effective richer set features respect standard convolutional neural network. Numerical experiments confirm model yields considerably better than state-of-the-art real-world problem