GLAD: Global–Local Approach; Disentanglement Learning for Financial Market Prediction

Figure 1

Overview of GLAD architecture. The model is made up of the following components: (a) an informer encoder to capture the temporal relationship between historical data, (b) a global feature disentangler that uses a mixture of autoregressive experts, (c) a local feature disentangler that acts in frequency domain, (d) time-based augmentation is used to augment the global feature by jitter, shift, and scale, (e) frequency-based augmentation by adding frequency, (f) time contrastive loss is used as discriminative global learning, (g) frequency contrastive loss which is used to discriminate local learning, and (h) an informant decoder which is used for final prediction.