5. PROFILE AND TRENDING ANALYSIS › 5.2 Usage Guidelines

5.2 Usage Guidelines

The principal items to consider when you use Profile and
Trending Analysis are the number of future intervals you wish
to estimate and whether you want the future estimates based
on the long- or short-term rate of change.

The number of intervals that can be forecast based on a
historical data series is a function of two variables:

o  The length (that is, number of observations) of the
   historical data series

o  The behavior of the historical data series

As a general rule, the number of forecast intervals that you
specify should never exceed the length of the historical data
series.  Limiting the length of the forecast to a maximum of
one-third or one-half the length of the historical data
series is often a far more judicious choice.

You must also consider the qualitative behavior of the
historical data series when you select the number of
intervals to be forecast.  If the historical data series
appears to represent a smooth trend, then you can specify a
forecast period one-half to two-thirds the length of the
data.  If the historical data series exhibits high
variability or apparent randomness, then you should specify a
very short period.

Quantitative techniques for determining the behavior of the
historical data series and the length of forecast are
included in Chapter 7, Univariate Model Forecasting.

You must also specify whether the future estimates will be
based on the short- or long-term growth rates.  Each of these
two options have both advantages and disadvantages.

o  If you specify the short-term growth rate, the future
   estimates are based on the rate of change exhibited
   between the final two observations.  You can use this
   option, which is extremely sensitive to the system's most
   recent behavior, to identify changes in long-term behavior
   patterns (that is, when the signs of the short- and
   long-term growth rates are different).

o  The short-term growth rate can produce deceptive results
   for data elements that you report as totals, such as total
   TSO LOGONs, when the MONTH level of detail is specified.
   This problem is caused by the varying number of days that
   is represented by any month.  Whenever possible, convert
   such variables to percents or rates to avoid the problems
   that are introduced by dissimilar measurement intervals.

o  The long-term growth rate produces much more conservative
   estimates, since the significant variations that can occur
   between any two time periods are dampened when more
   intermediate time periods are included.  This dampening
   results since the long-term rate of growth is the n-th
   root (see Equation 1 in Section 5.4) of the growth rate
   between the two observations.

For extended forecasts, the long-term growth rate often
represents a better choice, since the likelihood of
significant growth rates is small.  However, be careful to
avoid the long-term growth rate if there was some significant
change to the hardware or software that produced one or more
steps in the historical data series.