Скачать презентацию Short-Term Processes Forecasting by Analogues Complexing GMDH Algorithm

a39ceeb28e5d84e54889fe2e5ff33361.ppt

• Количество слайдов: 20

Short-Term Processes Forecasting by Analogues Complexing GMDH Algorithm Gregory Ivahnenko September 2008

Agenda Overview Description of the Analogues Complexing GMDH algorithm Example of Application Forecasting of the products launch Software system Conclusions

Basic points The Analogues Complexing algorithm was proven to be effective forecasting and clusterization of data samples in many applications. Solution of modern data mining problems require consecutive application of different inductive and deductive methods.

Spectrum of GMDH Algorithms

The Analogues Complexing (AC) algorithm Used for: - Classification, - Clusterization, - Stepwise forecasting of multidimensional random processes by complexing (weighted addition) of analogues (similar patterns) taken from historical data. The main parameters of algorithm are optimized by sorting of discrete number of possible variants by inductive algorithm. The main assumptions here are following: Investigated object can be described by multidimensional process; Multidimensional process is sufficiently representative, i. e. essential system variables are included into the input data sample and it contain sufficient number of observations; Part of previous behaviour of system in past can be repeated

Analogues Complexing (AC) GMDH Algorithm Forecast is calculated by complexing of optimal number of analogues. Where: A 0 - output pattern; Ai - analogues; Ai. F - predictions; F – number of predictions.

Analogues Complexing In rigid complexing of F predictions, the output prediction pattern A 0 F is defined using weights λi of analogues complexing (1) (2) where l 0 i – Euclid distance between initial pattern and analogues; F – number of predictions.

Steps of sorting The general problem of AC algorithm parameters optimization can be solved in four steps of sorting, in four dimensional space: 1. Set of input variables X; 2. Number of analogues F for complexing; 3. Length of analogues k; 4. Weight coefficients λi values for analogues complexing.

Application - Input data sources and structure Sales data for each product and store from all retailers. Hierarchical data for categories of products; Geographic data from customers postcodes; Demographic data from surveys and statistics; Media advertising data about corresponding TV, press and radio campaigns.

Media spend Brand X – Air freshener GB / Major Mults The impact of advertising can transform a new product, in certain categories, but typically lifts short term sales by 25% Weeks from launch

Benchmarks No Deal Volume Sales Rate Index No Deal Sales Rate Index Latest Quad-Week: Period-On-Period Change: Position: 22 -6 points

Example of Forecasting by AC algorithm

Analytic software system Main focus on knowledge discovery than on reporting. Was designed as a fully automated interface for analytic functionality to produce immediate insights in a reports using: Long-term forecasting for product launches. Scenarios’ simulation using marketing mix models. Media analysis. Flexible automated division of products and clusters analysis. Decision support helps to point out the most important variables for launch or to find features for the future product. The deck of slides is generated directly from the Excel system. Robust data import and export of products clusters.

Structure of Software System

Thank you for your attention. . .

Short-Term Processes Forecasting by Analogues Complexing GMDH Algorithm Gregory Ivahnenko [email protected] net

Features of the Modified Combinatorial GMDH algorithm Recurrent bordering method for parameter estimation Goedel’s number renumeration for restoring of the models structure at the end of computation Similar to Combinatorial- selective algorithm of V. S. Stepashko Ridge regression method

Current marketing problems Price elasticities and price thresholds analysis Marketing mix method, based on regression analysis investigate dependencies of profit from small deflections of price and find price thresholds, after which it will change significantly. Prediction of the new product launches for very small number of observations Long-term forecasting Analysis of relationships of sales from promotions and shoppers demographic data Media analysis optimize media campaigns to maximize profit. “Transferable demand” approach investigate any possible changes of sales or promotion strategy Clusterization of shopper, store and product characteristics Shopper surveys analyze demographics, attributes of products and shoppers paths in stores Segmentation algorithms select the most demanded attributes of a new products

Fragmented consumers means more precision Forecasted top 10 growth areas over the next 10 years Experian’s Mosaic UK Forecast data 2006 - 2016 Analysis based on forecasted % change in numbers of households

– Cura-Heat New Brand/Major Multiples 4 -Week Periods from launch BACK TO APPENDIX PAGE Category: Topical Analgesics Launch Type: Best Launched February 04