Adding Value to your Trend with Confidence Intervals - continued

To continue where I left off on Monday, let’s see what happens as time goes forward and more measurements become available.

Here’s what happened after the second month of measurements.  Clearly the upward trend is continuing – and look how much closer together the confidence limits are.  While you would not yet bet the farm on what will happen on 1st December, you can now see that the extremes of 35% (on the low side) and 90% (on the high side) are both extremely unlikely – indeed, you can be 95% confident that your measurements show a genuine trend between 50% and 70%, much narrower than previously.

However, through all this, please don’t forget Lesson 1 of my previous set of blogs – Most trends don’t extrapolate for ever – try to work out what the limiting factors may be.  In this example, the limit is probably a straightforward factor such as the number of new users still to be migrated to this application. When all the new users are happily set up, the trend will almost certainly stop rising – or at the very least, it will change significantly.  An external factor like this shows that confidence limits on a trend are only useful if the trend remains more or less constant.

In this case, there’s every expectation that the trend will remain more or less constant.  Here's the trend, with the same confidence limits, predicting utilization between 50 and 60%, on three months worth of data collection.
This looks like a trend we can rely on and certainly gives us more confidence than old fashioned guesswork.

In real life you will probably want to track the peak hour as well as the average hour.  That’s fine – you will end up with a range of confidence for the predicted utilization during the peak hour, and almost certainly a slightly different range of confidence for the predicted utilization during an average hour.

In future blogs I’ll look at some more complicated trending issues, including cases where trending just won’t work, and where you will have to use a different prediction technique such as modeling.

Andy Mardo
Product Manager