Model Stacking for TPV Prediction Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Accurately modeling and predicting Total Payment Volume (TPV) is challenging due to the complexity of data sources and modeling architectures, requiring improved methods to leverage various data effectively and detect anomalies for timely remediation in network and system environments.
Innovation Solution
The implementation of model stacking systems and methods that combine forecasting models with machine learning models, using weighted seasonality parameters and prediction enhancing data such as macroeconomic and internal metrics to generate accurate TPV predictions, allowing for real-time monitoring and automated remediation actions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If simple linear projection techniques are used for modeling, then the modeling process is simple and quick, but the prediction accuracy is insufficient
Solution Approach 1:
The patent combines multiple machine learning models (prophet model, LSTM model, and gradient boosting model) into a stacked ensemble architecture. Each model captures different patterns in the data, and their predictions are combined through weighted averaging to achieve superior prediction accuracy compared to individual models or simple linear projections.
Solution Approach 2:
The patent creates a composite modeling approach by integrating multiple heterogeneous machine learning models with different architectural strengths. The stacked ensemble combines time-series forecasting capabilities (prophet, LSTM) with tree-based regression (gradient boosting), creating a composite system that leverages the strengths of each model type to improve overall prediction accuracy.
2Measurement precision
If complex machine learning models are used to improve prediction accuracy, then prediction accuracy improves, but experimentation and development resources are required
Solution Approach 1:
The patent segments the complex modeling task into distinct components: a prophet model for trend and seasonality capture, an LSTM model for sequential pattern recognition, and a gradient boosting model for non-linear relationships. Each segment is independently trained and optimized, then their predictions are combined. This segmentation reduces the overall complexity by breaking down the monolithic complex model into manageable, specialized sub-models.
Solution Approach 2:
The patent introduces an intermediary layer (the stacked ensemble combination mechanism) that integrates predictions from multiple specialized models. This intermediary layer uses weighted averaging to combine individual model outputs, serving as a mediator that synthesizes the strengths of each model while managing the complexity of integrating multiple machine learning approaches.
3Measurement precision
If multiple data sources are integrated to improve TPV prediction, then prediction accuracy improves, but the complexity of leveraging data effectively increases
Solution Approach 1:
The patent applies local quality by allowing each machine learning model to focus on specific aspects of the data with appropriate weighting. The prophet model emphasizes temporal patterns and seasonality, the LSTM model focuses on sequential dependencies, and the gradient boosting model captures non-linear relationships. Each model processes the same multi-source data but with different local optimizations, reducing the complexity of integrating diverse data sources into a single monolithic model.
Data Source
AI summary
This specification includes machine learning model stacking techniques allowing for greater predictive accuracy using disparate sources of data. In one embodiment, a system obtains TPV data and inputs the TPV data into a forecasting model. Based on the total payment volume data, the forecasting model may output a first prediction of a total payment volume for a future period of time. The system may acquire prediction enhancing data and input the first prediction from the forecasting model and the acquired prediction enhancing data into a machine learning model. Based on the first prediction and the acquired prediction enhancing data, the machine learning model may output a second prediction of the total payment volume for the future period of time. The second prediction may be compared against real-time TPV and determined differences may be used for controlling operations of various machines system/network environment machines.