Time Series Forecasting via Data Routing and Algorithm Ensembling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional prediction approaches struggle with accurately forecasting demand for infrequently sold items with intermittent and seasonal patterns, often resulting in inaccurate non-zero demand predictions, especially for large lead times, which can lead to stockouts or overstocking issues in inventory management.
Innovation Solution
A machine learning-based forecasting service that employs distributed optimization algorithms and a flexible software framework for blended global-local optimization, allowing for item-specific and shared parameter optimization, and combines predictions from multiple learning algorithms to generate probabilistic forecasts, effectively handling heterogeneous demand characteristics and varying data lengths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional prediction approaches are used for infrequently sold items, then the forecasting process is simple, but the prediction accuracy for non-zero demands deteriorates
Solution Approach 1:
The patent segments the forecasting system into multiple specialized learning algorithms, each trained on specific subsets of time series data with different demand characteristics. This allows the system to handle intermittent, seasonal, and continuous demand patterns separately, improving prediction accuracy for infrequently sold items without requiring a single overly complex model.
Solution Approach 2:
The patent applies local quality by training different learning algorithms on locally relevant data subsets specific to each algorithm's strengths. Each algorithm receives tailored training data that matches its capabilities, enabling it to make accurate predictions for specific types of demand patterns while maintaining overall system manageability.
2Measurement precision
If multiple learning algorithms are combined for prediction ensembling, then the forecasting accuracy improves, but the computational complexity increases
Solution Approach 1:
The patent divides the computational workload by segmenting data routing to different learning algorithms based on time series characteristics. Each algorithm processes only the subset of data it is best suited for, avoiding the computational overhead of running all algorithms on all data while still achieving ensemble accuracy.
Solution Approach 2:
The patent dynamically changes parameters such as routing decisions and algorithm selection based on time series properties. This allows the system to adapt computational resource allocation to the specific characteristics of each forecasting request, optimizing the balance between accuracy and computational efficiency.
3Measurement precision
If data is routed to specialized learning algorithms, then the prediction accuracy for specific demand patterns improves, but the system complexity increases
Solution Approach 1:
The patent introduces a data routing layer as an intermediary between the time series input and the multiple learning algorithms. This routing mechanism automatically directs appropriate data subsets to the most suitable algorithms based on predefined criteria, improving prediction accuracy while shielding the user from the underlying system complexity.
Solution Approach 2:
The patent creates a universal forecasting system that can handle multiple types of demand patterns through a single multi-functional architecture. The system universally processes various time series by routing them to appropriate specialized algorithms, providing accurate forecasts for diverse demand characteristics without requiring separate systems for each pattern type.
Data Source
AI summary
An input data set with a plurality of item descriptors comprising respective time series observations is identified. A routing directive indicating a predicate to be evaluated to determine whether a particular item descriptor is to be included in a training data set for a first learning algorithm is obtained. A plurality of learning algorithms are trained using training data sets derived from the input data set according to respective routing directives, and the trained algorithms are stored.