Demand Forecasting With Dynamic Market Share Decomposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing parametric methods struggle with time-series and auto-correlated effects in demand forecasting, while deep learning techniques inaccurately represent unobserved arrival rates using static data, and both fail to capture complex nonlinear relationships in censored data scenarios.
Innovation Solution
An automated forecasting system using Mixed Integer Programming (MIP) and Temporal Convolution Networks (TCN) decomposes censored data into time-dependent arrival rates and purchase probabilities, enabling accurate forecasting of future demands and market shares by training a neural network with simultaneous first and second temporal networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If parametric methods like Multinomial Logit are used for demand forecasting, then the model is simple and easy to implement, but it cannot capture time-series effects, auto-correlation, or complex nonlinear relationships
Solution Approach 1:
The patent transitions from simple parametric models to deep learning models with dynamic parameters that can capture time-series effects and nonlinear relationships. The neural network learns complex patterns in the data while maintaining computational efficiency through parameter sharing and pre-computation of arrival rates.
2Measurement precision
If deep learning methods are used to handle censored data, then time-dependent patterns can be captured, but the use of pre-calculated static arrival rates fails to represent dynamic unobserved arrival rate patterns
Solution Approach 1:
The patent makes the arrival rate dynamic by computing it as a function of time and contextual features using a neural network. This allows the model to adapt to non-stationary patterns and time-dependent variations in demand, rather than relying on static pre-calculated rates.
3Ease of operation
If existing techniques are used on stationary past data, then the models work adequately, but they fail when data is non-stationary and contains unobserved lost shares
Solution Approach 1:
The patent introduces an intermediary component that estimates unobserved arrival rates and lost shares using neural networks. This intermediary layer bridges the gap between observed purchase data and the underlying demand processes, enabling reliable forecasting even with censored data.
4Measurement precision
If simultaneous training of multiple temporal networks is implemented, then forecasting accuracy improves, but computational time and resource requirements increase
Solution Approach 1:
The patent segments the forecasting task into multiple specialized temporal networks, each handling different aspects of the prediction problem. This segmentation allows for more accurate modeling of complex patterns while enabling parallel training and efficient resource utilization.
Data Source
AI summary
In an aspect, input data can be received, including at least time series data associated with purchases of at least one product and causal influencer data associated with the purchases. The causal influencer data can include at least non-stationary data, where lost shares associated with said at least one product are unobserved. An artificial neural network can be trained based on the received input data to predict a future global demand associated with at least one product and individual market shares associated with at least one product. The artificial neural network can include at least a first temporal network to predict the global demand and a second temporal network to predict each of the individual market shares. The first temporal network and the second temporal network can be trained simultaneously.