Masked Multi-Step Forecasting for Time Series Accuracy

The patent segments the forecasting task into multiple independent direct prediction steps, where each step predicts a specific future time step directly from historical data without relying on previous predictions. This segmentation prevents error accumulation by eliminating the recursive dependency chain that causes errors to propagate and amplify over longer forecasting horizons.

The patent introduces an intermediary training mechanism using masked future information during the training phase. By masking certain future time steps and training the model to predict them, the system creates an intermediate learning stage that improves the model's ability to make accurate direct predictions without the harmful effects of recursion during actual forecasting.

The patent applies preliminary action by performing masked training on future information before actual forecasting. During training, certain future time steps are masked and the model learns to predict them using historical data and unmasked future information. This preliminary training phase prepares the model to make accurate direct predictions during inference without requiring complex training procedures, as the model has already learned the prediction task in advance.

The patent implements feedback mechanisms during training by comparing model predictions with actual masked future values. This feedback loop allows the model to learn from its prediction errors and continuously improve its direct prediction capability, making training more effective and less difficult even for large forecast horizons.

The patent extracts and utilizes only the necessary future information (predictor variables) that are needed for improved predictions, while keeping the forecasting framework simple. By selectively taking out and using relevant future predictor variables without incorporating entire future sequences, the system achieves accuracy improvements without proportionally increasing framework complexity.

The patent creates a universal forecasting framework that can handle both traditional recursive forecasting and the new direct forecasting with future information using the same model architecture. This multi-functional approach allows the system to incorporate future information when available while maintaining simplicity, as the same framework serves multiple forecasting purposes without requiring separate complex systems.