Demand Forecasting and Capacity Planning Overview:

The following offers a summary of our expertise in providing world-class demand forecasting and capacity planning. We present our unique approach to demand forecasting for the telecommunications market.

Datamining Approach to Forecasting Telecommunications Demands and Network Loads:

Many telecommunications companies are placing an increasingly high economic value on the ability to accurately forecast demand for local loop and access products, ranging from primary residential lines (1FRs) to wireless, data, and bundled local/long-distance services, over a planning horizon of several years. Failure to forecast demand growth with sufficient accuracy has recently led many experienced providers to inadequate capacity provisioning, held orders, and sometimes regulatory penalties or churn of customers away to competitors.

In response to the need for more accurate demand and network load forecasts, Cox Associates has developed and successfully applied a new approach to demand forecasting that can radically improve forecast accuracy compared to even the most sophisticated time series methods.

The new technique, which we call Classification-Tree Compartmental Forecasting, is based on the following simple steps, implemented with the help of sophisticated data-mining and stochastic simulation algorithms:

Several features make classification trees particularly well suited to avoid many of the threats to valid causal inference identified in Table 1. Specifically:

	Divide units of demand- or load-analysis (households, firms, subscribers, base stations, etc.) into groups with relatively homogeneous demand behaviors, e.g., for acquisition, usage, and churn or attrition of local access products.
	Forecast the behavior of each group. This can be done accurately, by construction.
	Sum forecasts over all groups to obtain aggregate forecasts.

1. The almost infinite number of ways to partition the units of analysis into groups. Identifying groupings that minimize forecast errors has been a continuing challenge for marketers and statisticians.

2. Difficulties in predicting the flow of units among the different groups over time.

Our technique provides constructive solutions to both problems, as follows.

1. Individuals are automatically partitioned into groups (defined by conjunctions of attribute values) to minimize forecast errors. This partitioning is accomplished via the data-mining and artificial intelligence technique of classification tree analysis. The quantities forecast -- product acquisition, penetration, usage, and retention/attrition/churn rates -- are modeled as transition rates in a multi-state, semi-Markov process with covariates that include demographic, billing, and usage history information. Transition rates are estimated as functions of these covariates via the classification tree algorithm.

2. Transitions of individual units among groups are modeled by applying the transition rates estimated in part (a) to the groups. This determines the evolution of the distribution of individuals among groups over time. A dynamic simulation model is used to integrate the transition rate and covariate information and to predict the resulting changes in product demands over time.

Cox, L.A., Jr. "Using classification trees to improve causal inferences in observational studies". Preliminary Papers of the Sixth International Workshop on Artificial Intelligence and Statistics. 1997.

Cox, LA, Jr., "Using causal knowledge to learn more useful decision rules from data," Chapter 2 in D. Fisher and H.-J. Lenz (eds), Learning from Data: AI and Statistics V. Springer-Verlag, 1996.

Cox, L.A., Jr., G. Bell, and F. Glover, "A new learning approach to process improvement in a telecommunications company." Production and Operations Management, 4, 3, 217-227, 1995.

Cox, L.A., Jr., "Combining the probability judgements of experts: Statistical and artificial intelligence approaches", Chapter 26 in D.J. Hand (Ed), Artificial Intelligence Frontiers in Statistics. Chapman and Hall, 1993.

Cox, L.A., Jr., Y. Qiu, and L. Davis, "Guess-and-verify heuristics for reducing uncertainties in expert classification systems," in D. Dubois et al (eds), Uncertainty in Artificial Intelligence. Morgan Kaufmann, San Mateo, CA, 1992.

Cox, L.A., Jr., "Pragmatic information-seeking strategies in expert classification systems," in D. Brown and C. White (eds), Operations Research and Artificial Intelligence: The Integration of Problem-Solving Strategies. Kluwer, New York, 1990.

Cox, L.A., Jr., "Incorporating statistical information into expert classification systems to reduce classification costs," Annals of Mathematics and Artificial Intelligence, 2, 93-108, 1990.

Cox, L.A., Jr., "Managing uncertain risks through 'intelligent' classification: A combined artificial intelligence/ decision analysis approach," pp 473-482 in J.J. Bonin and D.E. Stevenson (eds), Risk Assessment in Setting National Priorities. Plenum Press, New York, 1989.

Cox, L.A., Jr., Y. Qiu, and W. Kuehner, "Heuristic least-cost computation of discrete classification functions with uncertain argument values," Annals of Operations Research, 21, 1-30, 1989.

Find Out More About How Classification Trees Work!

Send mail to webmaster@cox-associates.com with questions or comments about this web site.
Copyright © 2001 Cox Associates, Incorporated. All rights reserved.