Forecasting System Using Lagged Correlation and Segmented Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face inefficiencies and inaccuracies in predicting future values for multiple products due to the need for independent training of machine learning models and difficulties in implementing global forecasting models across large datasets, often resulting in mediocre accuracy across all products.
Innovation Solution
A system that retrieves time-series data for each product, computes correlation values with other products at various degrees of lag, selects a subset of products based on correlation thresholds, and uses a machine learning model to predict future values, with the ability to choose between LASSO, Random Forests, or Deep Learning depending on data dimensions, and interpolates missing data to synchronize temporal frequencies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If machine learning models are trained independently for each product, then model training is simple and straightforward, but training efficiency is low and prediction accuracy is reduced
Solution Approach 1:
The patent segments the training process into two distinct phases: (1) a global model trained on aggregated data from all products to capture cross-product patterns, and (2) product-specific models trained on individual product data to capture unique characteristics. This segmentation allows the system to benefit from both efficient global learning and accurate local prediction, resolving the contradiction between training simplicity and efficiency.
Solution Approach 2:
The patent merges the global model predictions with product-specific model predictions through a weighted combination approach. The global model captures general trends across all products efficiently, while product-specific models refine predictions for individual products. This merging strategy achieves both high training efficiency (from the global model) and high prediction accuracy (from the product-specific models).
2Adaptability or versatility
If a global forecasting model is built across all products, then cross-product interactions are captured, but prediction accuracy for individual products becomes mediocre
Solution Approach 1:
The patent segments the forecasting approach into a global component and local components. The global model captures cross-product interactions and general patterns, while separate product-specific models capture individual product characteristics. This segmentation allows the system to maintain both broad adaptability (from the global model) and precise individual predictions (from product-specific models).
Solution Approach 2:
The patent applies local quality by allowing different modeling approaches and parameters for different products based on their specific characteristics. Each product can have its own model architecture, hyperparameters, and data preprocessing steps tailored to its unique patterns, while still benefiting from the global model's cross-product insights. This ensures high prediction accuracy for each individual product.
3Measurement precision
If existing systems leverage interactions between products, then forecasting accuracy may improve, but implementation becomes difficult on datasets with large number of products
Solution Approach 1:
The patent segments the complex multi-product interaction problem into manageable components: a global model that handles cross-product interactions in a unified framework, and product-specific models that handle individual product nuances. This segmentation makes the implementation scalable to large numbers of products by avoiding the need to model all possible product interactions explicitly.
Solution Approach 2:
The global model serves as an intermediary that captures cross-product interactions and patterns, which then inform the product-specific models. This intermediary approach simplifies implementation by providing a standardized framework for handling product interactions, rather than requiring complex custom implementations for each product pair.
Data Source
AI summary
Systems, software, and methods are disclosed for generating a prediction of time-series data from data sets. A system is configured to: retrieve, for each product of a set of products, the time-series data including time-value pairs; select a first product; compute a correlation value between the first product and other products and for one or more degrees of lag to obtain a set of correlation values representing the correlations between the first product to the other products assessed at prior times; select a subset of products based at least in part on the correlation values; provide the time-series data associated with each product from the subset of products and the first product to a machine learning model trained to predict a future value of the first product based on values of the subset of products; and obtain prediction data representing a set of predicted values for the first product.