Trends and Forecasts

To reduce the impact of cyclic changes in the historical data, a user-specified business period is used to break the data into time-interval based bins and each bin is then represented by a single point. The point can be the average, the peak, or the 90th percentile of the data (90% of the points are less than the value). A bin will not be used unless the percent of points within the bin that are valid exceeds the threshold specified by the user.

	IMPORTANT: A trend will not be calculated unless at least two bins with an adequate percentage of valid points exist within the range of data being analyzed.

A user can set the report period to exclude a special event or mark the time period invalid to exclude points collected during that period from a trend analysis.

The linear regression is based on a least squares fit that minimizes the sum of the squares of the vertical offsets between each of the aggregate points and the trend line that describes them.

	TIP: Regressions performed over small data sets are not always meaningful and can be misleading. Any trend analysis based on less than a dozen aggregate points should be carefully compared with the historical data to see if it "makes sense." The maximum number of data points for the trend analysis is the total time for the report divided by the business period, since business periods can be excluded if they do not meet the validity criteria. Since the trend is reported as an annual growth rate, it is best to have more than a year of historical data before trying to analyze trends.

The user can choose to include error analysis in the report. The following error value is available:

r-squared: The correlation coefficient of the regression performed to estimate the trend. A measure of how well the regression represents the aggregate points. Expressed as a decimal value; values approaching 1 or -1 indicate the regression more accurately represents the data.

The forecast model can be specified at four different levels within Capacity Advisor, with more specific forecast models overriding more general models, as indicated in the following table:

The forecast data range defines the historical data that is combined with the annual projected growth rate to produce the forecast model. The forecast data range can be specified as:

The annual projected growth rate is specified in percent and can be positive for increasing utilization, negative for decreasing utilization or zero for no change. No change is the default. Separate rates can be specified for memory and CPU growth.

The forecast is applied point-by-point to the historical data within the range specified by the user. It is applied linearly, so that a point 1 year from the starting point of a forecast is the result of the full growth rate being applied to the data. The data within the range provided by the user is used to “tile” the future by applying the portion of the growth rate appropriate to each point to each point in the data range and repeating the data set until the desired end point is reached.