Time Series Forecasting Models for Sparse Hierarchical Data

The patent transforms univariate time series data into multivariate representations by creating multiple derived time series from a single source series. This includes generating reversed time series, differentiated time series (first, second, and third derivatives), and integrated time series. By transforming data from one dimension (single time series) to multiple dimensions (multiple related time series), the system creates sufficient training data volume for deep learning models without requiring additional external data sources.

The patent segments a single time series into multiple distinct time series components, each capturing different characteristics of the original data. The segmentation creates: original time series, reversed time series (capturing backward patterns), differentiated time series (capturing rates of change at multiple orders), and integrated time series (capturing cumulative patterns). This segmentation allows the deep learning model to learn from diverse patterns within the same underlying data source.

The patent addresses hierarchical relationships by transforming hierarchical time series data into a multivariate format that deep learning models can process efficiently. By creating multiple derived time series from hierarchical entities (including parent-child relationships in entity hierarchies), the system maintains the hierarchical information while presenting it in a flattened, processable format. This approach captures hierarchical patterns without requiring complex hierarchical processing architectures.

The patent introduces derived time series as intermediary representations that bridge the gap between hierarchical data structures and deep learning model inputs. These intermediaries (reversed, differentiated, and integrated time series) serve as mediators that encode hierarchical relationships in a form suitable for standard deep learning architectures, avoiding the need for specialized hierarchical processing while preserving hierarchical information.

The patent applies data transformation processes (flattening and stacking) as preliminary steps before model training and deployment. By pre-processing the time series data into the required multivariate format during the data preparation phase rather than during real-time inference, the system incurs the computational cost only once during setup. The transformed data structure is then reused for training and forecasting operations, minimizing repeated processing overhead.

The patent segments the data processing workflow into distinct transformation steps (flattening, stacking, reversing, differentiating, integrating) that can be independently optimized and executed. This segmentation allows for efficient implementation where each transformation operation is performed once on the raw data to create the complete set of derived time series, which are then cached and reused for multiple model training iterations and forecasting operations.