Annotating Data Models for Predictive Systems

The patent extracts and separates anomalous data points from the historical dataset before performing predictive modeling. The system identifies outliers using statistical methods (e.g., IQR, z-score) and removes them from the training data, allowing the model to learn only from normal usage patterns. This resolves the contradiction by eliminating the harmful anomaly skew while preserving the beneficial historical pattern learning.

The system implements a feedback mechanism where predicted values are compared against actual values, and anomalies in the residuals are identified and fed back into the modeling process. This iterative approach allows the system to continuously refine its understanding of normal patterns by learning from prediction errors, thereby improving predictive accuracy while accounting for anomalies.

The patent applies data filtering techniques to extract and remove anomalous observations from the historical dataset. Methods such as statistical outlier detection, moving average filtering, and seasonal decomposition are used to identify and exclude distorted data points. This allows the model to utilize the full volume of historical data while eliminating the portions that would skew the forecasts.

The system applies different quality standards to different portions of the historical data. Normal data points are included with full weight in the training process, while anomalous points are either removed or down-weighted. This local differentiation in data quality treatment allows the model to learn from the abundant normal data while being insensitive to the harmful effects of anomalies.