Info

Notes: Experiments are described in the text. The CUSUM results are the best results out of CUSUM runs with all values of the H parameter between 1 and 100. A more complete version is provided in Table 14.2.

Notes: Experiments are described in the text. The CUSUM results are the best results out of CUSUM runs with all values of the H parameter between 1 and 100. A more complete version is provided in Table 14.2.

• What fraction of spike outbreaks are detected if alarm thresholds are set at a level that produces one false alarm every ten weeks? This performance must be inferior (i.e., a lower fraction detected) to the two-week case because the alarm threshold must be higher.

• How many days on average does it take until a ramp outbreak is detected if alarm thresholds are set at a level that produces one false alarm every two weeks? If the ramp is undetected we score it as requiring five days.

• How many days on average does it take until a ramp outbreak is detected if alarm thresholds are set at a level that produces one false alarm every 10 weeks?

From the table we see that simple control charts perform poorly (because of seasonal problems, and because of serious day-of-week effects). When we use "yesterday" we do not suffer from the seasonality-induced problems but day-of-week effects are even worse. Moving average is more effective (although methods described later in this chapter will thoroughly resolve it), and CUSUM is, on these examples, relatively insensitive.The performance of MA seems to be little affected by the window size W: it is best at seven days and degrades only slightly if it is halved or doubled.

Was this article helpful?

0 0

Post a comment