|
Predictive Modeling
Creating business intelligence from subscriber usage information
By: Bruce Bahlmann - Contributing Author (your
feedback
is important to us!)
Created: August
24, 2002
| Published by: |
CED
|
-- |
December 2002 |
Note: For help applying predictive modeling techniques to your system or developing tools to help you audit, test, and track the success of your efforts contact Birds-Eye.Net.
Predictive modeling is a means of looking at current events through special lenses
(which are shaped and defined by the analysis of historical data) to identify patterns as
well as predict future events. Once thought of as science fiction, predictive modeling is
now ranked among the top ten innovations changing today’s information technology
landscape. This article will examine some of the areas where predictive modeling is
showing up within broadband, compare these areas with existing tools being offered, and
finally look at suggested implementation. This article discusses/compares the benefits and
capabilities of predictive modeling rather than traversing its associated mathematics and
analysis.
Past vs. Present
Have you ever heard of the phrase, “history repeats itself?” Well, this
phrase is so well known and understood that it is essentially a given - a fact of life.
Yet most products and services that we buy today evolve around the present. Products that
deal with the present capture most people’s interest over those that merely deal with
the past or a combination of something and the past. Perhaps the “past” has just
got a bad rap. It is as if our hurried lifestyle has no time for the past and all we want
is current events. Up to the minute news in 60 seconds is popular, not 60 seconds of
significant historical patterns.
In order to determine the possible applications of predictive modeling within the
broadband space it is helpful to understand how Broadband Service Providers (BSPs) view
the importance of past vs. present. Take for example how BSPs approach the management of
their networks. The number one thing that BSPs want is the status of a device. Is it up or
is it down, current power levels, bandwidth usage, etc. rather than usable shelf life,
service history, rate of capacity exhaustion, predicted expiration date. If BSPs had their
choice of monitoring status of the device or having historical information on that device
all would rather have the current status (up to the minute is preferred). Past events are
though of as interesting but not mission critical. Not to take anything away from status
– it can be important (perhaps the most critical information depending on the
device). However, depending solely on the current status of things forces you to be much
more reaction oriented. Current events also become increasingly meaningless without past
history to provide some basis on which to understand them. Gathering current status is
also extremely resource intensive that may well be limiting from a network bandwidth usage
perspective.
For example, if the status of a subscriber device goes down, this would produce
different responses based on whether you’re interested in present events or past
historical patterns. If you’re mainly looking at present information this event may
require you to react (perhaps roll a truck) or alternatively check additional devices
within the same geographic vicinity for similar behavior. However, if this event were
examined using historical patterns it may merely indicate the subscriber’s normal
usage pattern and either get ignored or further reinforce the current model. BSP address
the limited capabilities (i.e. false alarms) of status monitoring by looking at large
numbers of devices often assembling them into groups in addition to examining individual
devices. If a number of devices register some kind of alarming behavior this would more
likely indicate some type of problem – but even this is not known for sure. Examining
individual devices or even geographically close devices is at best an indicator – one
of many such data points available in determining network health and subscriber activity.
In reality, a large number of such data points exist that together explain the overall
operation of devices on a network. Engaging in a practice of relying on current status can
thus lure the BSP into becoming increasingly reactive and overly status monitoring
inquisitive (poll happy). However, combining status monitoring with predictive modeling
can reduce the amount of information required of a device to evaluate the severity of
events. Note that predictive modeling does not replace status monitoring but rather makes
it more efficient and accurate.
It is worth noting that some of the carrier class Network Management System (NMS)
software packages have begun offering more predictive modeling like capabilities to BSPs
using the data they already collect. Having status information in hand without duplicate
polling makes these implementations attractive as these initial product offerings
mature.
Predictive modeling vs. Bandwidth Management
Another area where predictive modeling goes head to head with existing technology is
with regards to bandwidth management. Note that inevitably, nothing can substitute for
increasing the usable capacity or efficiency of one’s network through the use of
software over upgrading or adding transmission equipment so the value of any interim
substitution is only the value of delayed expenditures. With that said, some tough
questions must be asked before one can evaluate the resulting benefit that either
bandwidth management or predictive modeling represent. Here are some questions I encourage
everyone to think about before seeking out bandwidth management applications: How much
time, effort, and money should be invested in fine-tuning existing resources before the
returns of such exercises begin to yield increasing smaller cost savings?
- How much is that value worth today and will that value continue to yield similar results
as networks become increasingly complex and diversified? (In other words is this a
short term investment or a long term investment and if long term what is the
software’s projected shelf life?)
- What is the shelf life of existing equipment and at what point does it actually cost you
less money to upgrade your network appliances over fine-tuning existing equipment (taking
into account capacity planning)?
- What is the likelihood of the capability demonstrated by a prospective software
application becoming embedded within the hardware it attempts to fine-tune? (In other
words does this application represent a short-term fix or a long-term solution/business?)
Predictive modeling as well as bandwidth management within this context should be
viewed as an expense whose aim is to create value by making existing equipment or
operations more efficient. Both represent tools that can gather various types of raw data
and then present them in such a way that they can convey the most economically meaningful
advice to their operator. However the results of these methods target different things:
Predictive modeling seeks perfection of its models where as bandwidth management seeks
perfection of the transmission system and its components. The benefit of having an
increasingly perfect model is that it one can proactively explain network and subscriber
behavior with increasingly fewer data points. The benefit of achieving an increasingly
perfect transmission system is that one can maximize the performance of existing system
hardware. It is not hard to see how bandwidth management becomes increasingly
reactive through rigorously changing the very system it seeks to perfect. While the same
could be said of predictive modeling during the modifications of its models, the result of
the model changes further evolve and substantiate the accuracy of the working model to
help make future decisions based on this model more reliable. Bandwidth management
claims to be able to roughly predict the future bandwidth demand by determining some type
of trend of current events and then extrapolating that trend out to some future date.
However bandwidth consumption is seasonal, changing, difficult to predict, yet ever
increasing so extrapolating this information is inherently flawed. In addition, bandwidth
management changes to the system can spoil previously gathered data and can further render
extrapolations useless. Predictive modeling exploits evolving subscriber usage models
along with current usage data to fairly accurately predict ongoing bandwidth usage.
Service Changes and Financials
In addition to the network and operations applications discussed, predictive modeling
is increasingly being used within billing systems. For example, various billing vendors
are currently offering predictive modeling packages for churn. BSPs who understand churn
know that effectively lowering churn requires more than just bundling services. It
requires an increasing list of customer care, billing, and operational related services so
having something keep track of the impact of each of these areas as well as each BSPs
effectiveness at reducing churn is attractive (see Figure 1.0).
Figure 1.0 Predictive Model for Churn
Implementing Predictive Modeling
As broadband service providers (BSPs) begin the process of migrating from flat rate
all-you-can-eat data service to a multiple tiered usage-based data service unique
opportunities to implement predictive modeling will be presented. The opportunity to
devise a data collection scheme for usage-based data service from the ground up will be
critical to successfully implementing predictive modeling in the future. The Record
Keeping System (RKS) seems to be a logical place to start looking for increasingly more
predictive modeling capability. While some vendors have already gone about adding this
capability to their RKS further BSP interest and requirements are needed to make this
capability the norm rather than the exception. Note that many newer services such as
PacketCable and CableHome leverage the RKS as part of their operation but predictive
modeling is not part of the specification so it is left to the vendor to build this on
their own perhaps in an attempt to differentiate their products from their competition.
It is conceivable that predictive modeling will not represent a unique application
or appliance running within your network but rather an important feature of various other
applications you use today. Therefore implementing this capability may only require you
upgrade certain applications to access this technology.
Conclusions/Recommendations
While history and knowledge of historical events mostly take a back seat to what is
currently happening, the significance and value of historical knowledge seems to
persevere. Like historical knowledge, predictive modeling too will persevere. However,
those BSPs who design their usage collection with predictive modeling in mind will be able
to leverage this data for more than just billing and network management. Predictive
modeling is the precursor to such things as asset tracking and financially linked network
assessment. Predictive modeling will help BSPs link the equipment costs they incur to
expand and maintain their networks to proposed revenue opportunities.
BSPs should exploit predictive modeling to track each of their assets like the airline
industry for example track their equipment. The airline industry classifies each airplane
as a valuable asset and thus tracks it independently. As a result, each airplane is
decommissioned, sold, etc. according to a schedule that takes into account mechanical and
operational history, performance, cost, and efficiency. The airplane is to the airline
industry as large network appliances are to the broadband industry – an instrument
that is necessary to do business. Large network appliances can cost BSPs up to a
half-million dollars or more but are not individually tracked in terms of cost. BSPs
likely waste a lot of money maintaining old or new but troubled equipment mainly because
expenses are not closely tracked along side the assets they represent. In the grand scheme
of things these expenses are pooled into a single bucket of operational costs but when
examined more closely, these costs could be broken down into maintenance costs of a
handful of troubled equipment – much of which could be avoided if that equipment were
replaced.
Don’t expect predictive modeling capabilities from vendors or applications that do
not represent authoritative data sources. The billing system, RKS, and NMS are natural
locations for predictive modeling capability because they maintain their own authoritative
data for their respective area. For example, don’t look for predictive modeling churn
capabilities from vendors other than your current billing software vendor. If you do,
don’t expect to see the same level accuracy as you would from your current billing
vendor. You may even have integration issues as billing systems don’t always expose
all their data which could limit some third parties from integrating predictive modeling
capabilities with your billing system.
Can Birds-Eye.Net help you or your Company?
Receive your Birds-Eye.Net articles and white
papers hot off
the presses by adding our RSS feed to your reader.
|