Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Method and System for Forecasting Future Inventory of Highly Seasonal and Event-Driven Inventory Traffic

IP.com Disclosure Number: IPCOM000201012D
Publication Date: 2010-Nov-04
Document File: 4 page(s) / 93K

Publishing Venue

The IP.com Prior Art Database

Related People

Konstantin Salomatin: INVENTOR [+4]

Abstract

A method and system for forecasting future inventory of highly seasonal and event-driven inventory traffic is disclosed. The method involves historical signal modeling (training) and subsequently forecasting the future inventory based on the historical signal model.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 52% of the total text.

Method and System for Forecasting Future Inventory of Highly Seasonal and Event-Driven Inventory Traffic

Abstract

A method and system for forecasting future inventory of highly seasonal and event-driven inventory traffic is disclosed.  The method involves historical signal modeling (training) and subsequently forecasting the future inventory based on the historical signal model.

Description

Disclosed is a method and system for forecasting future inventory of highly seasonal and event-driven inventory traffic.  The method and system involves historical signal modeling (training) and forecasting the future inventory based on the historical signal model.  The goal of historical signal modeling is to eliminate noise and uncertainty and also to learn global trends.  The signal is modeled as a Gaussian Process (GP) with a linear mean function and noisy Gaussian kernel as a covariance function.  The model hyper parameters (signal variance, scaling and signal noise) are learned by log-likelihood maximization in two iterative stages.  On the first stage, the objective function is optimized by a combination of analytical optimization and Limited memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) numerical optimization.  This gives an initial estimate of model parameters.  On the second stage, the signal noise estimate is refined using Leave One (year) Out Cross Validation (LOOCV) with forecasting error as an objective function.  The optimization is performed by line search.  The rest of the parameters (slope, intercept, signal variance and scaling) are adjusted by log likelihood maximization.

Figure 1 illustrates an example of a trained GP with mean function (black curve) and variances function (green curve).  The GP characterizes the signal with no global growth but zigzag patterns within events.

Figure 1

Figure 2 illustrates a signal with global trend.  Here, the GP has learned both the global trend and the event bumps in this signal.

Figure 2

In order to train the GP for a time serials, the following inputs are considered for a GP training algorithm:

1. : vector of input index values (N points, M features).  In the simplest case, there is only time index in X (M=1).

2. : vector of historical supply.  The value may be either the absolution number of impressions or the log value of the number of impressions.

3. : definition of a kernel function.

The training algorithm outputs a GP model, which consists of the following...