In this paper, a new modification called Bidirectional GLDM of the Generalized Least Deviation Method (GLDM) is proposed for robust and efficient forecasting of time series. With the inclusion of both forward and backward dependencies in its architecture, the proposed model enhances interpretability and predictive performance. By adopting a novel loss function based on the arctangent of residuals, Bidirectional GLDM offers resistance to outliers and captures complex interaction effects present in the data. Computational simulations applied to COVID-19 infection data in Moscow show that Bidirectional GLDM outperforms traditional models and the standard GLDM. Bidirectional GLDM significantly improves first-order metrics such as MAE, RMSE, and MAPE. Overfitting regularization and numerical difficulties remain for second-order models, indicating the need for stronger regularization. Results demonstrate the potential of Bidirectional GLDM in real-world applications requiring accurate time series prediction. The proposed method provides a robust basis for handling challenging datasets while maintaining both accuracy and interpretability by integrating bidirectional dynamics and leveraging interaction terms.