Finger flexion decoding and classification has gained attention to understand the relationship between finger movements and the brain activity. Initially focused on EEG signals, it was moved quickly to electrocorticography (ECoG) signals because of the advantages provided by the latter signals. The present paper proposes two based-CRF discriminative classifiers as approaches to the task: a CRF classifier and a Latent Dynamic CRF (LDCRF) model. Proposed classifiers have the advantage to take into account time dependencies without implementing generative models. Results show that proposed classifiers work better with high gamma (HG) band (accuracy: CRF: 0.742, LD-CRF: 0.737; Cohen's kappa: CRF: 0.682, LD-CRF: 0.681) than low-frequency components (LFC) (accuracy: CRF: 0.433, LDCRF: 0.430; Cohen's kappa: CRF: 0.322, LDCRF: 0.329). Compared with other classifiers and using HG data, based-CRF classifiers have higher performances (accuracy: CRF: 0.742, LDCRF: 0.737; Cohen's kappa: CRF: 0.674, LDCRF: 0.669) than three other classifiers: Linear Discriminant Analysis (accuracy: 0.224, Cohen's kappa: 0.024), Quadratic Discriminant Analysis (accuracy: 0.224, Cohen's kappa: 0.026), and a Linear Support Vector Machine (accuracy: 0.280, Cohen's kappa: -0.003), with a significance level of 0.05.