Calculating Exponential Function slope, using only partial data

[theblot]

Here's the situation: We're trying to project growth over a few years using our past data for reference. Simple exponential trendline, right? The problem is that in the last two weeks, we saw a massive jump which was out of average and will likely not happen again.

If I was using a linear trendline, it would be easy. I just calculate the slope using past data and apply it to the function. Is there a similar way to figure this out with an exponential tredline?

 

[Latexman]

Have you tried the Exponential trendline built into Excel's chart capability?

Good luck,
Latexman

 

[theblot]

Yes, but again, I need to exclude data from the slope, but include it into the final results.

Here's an example:

Let's say every week, I'm increasing exponentially by 1. So it goes as such:
1,2,4,8,16,32,etc...

But two weeks, it's off by a lot:
1,2,4,8,16,58,65, etc...

How can I figure out the growth slope, without considering the extraordinary data entries?

 

[electricpete]

Attached is a fit of an exponential function y = beta * exp(alpha*x) using solver. There is a weighting factor assigned to the square resiudal for each data point. The fit solution minimizes the sum of weighted square residual errors.

I put a zero in for the "suspect" data point. Whether it is really bad data or bad model is up to you to decide.

This type approach is also sensitive to the initial value guesses that you put into alpha and beta (like most solver problems).






=====================================
(2B)+(2B)' ?

 

[electricpete]

Correction
"I put a zero in for the suspect data point"
should've been:
"I put a zero [b[weighting factor [/b]in for the suspect data point"

=====================================
(2B)+(2B)' ?

 

[Denial]

How do you know which data to exclude? There is an entire sub-area of statistics devoted to the determination of "outliers".

 

[theblot]

This specific application is quite simple. We're calculating data storage on our network, to project how much storage we need to purchase for our company. Our data samples are from the backup routine that runs once a week. Well, for the past two weeks, we consolidated a massive amount of data from an archive source. We won't be doing that again (at least not for a long time), but that means the last two weeks are not a good representation of our normal growth rate.

 

[Latexman]

Then don't put that data in the regression set.

Good luck,
Latexman

 

[theblot]

Yea, here's what I did.

I constructed the "slope" based on prior data. Then, I added the difference between the last data point in the list and the data point not in the list. Therefore, it'll include the increase in data, without affecting the slope.

 

[MortenA]

Denial: WRT the subject of outliers, this is normally blisfully ignored

Best regards

Morten

 

[Denial]

Ah, blissful ignorance. You can't beat it.
(He who Grubbs around among the outliers often bites off more than he can chew.)

 

[theblot]

WRT?

 

[Latexman]

WITH RESPECT TO

Good luck,
Latexman

 

[GregLocock]

The easy way round is to prepare a second column of figures without the 'outlier' (or as we cynics might say, margin of error for your predictions).

Plot the true data as a curve using markers only, and the edited column the same. Add a trend line to the second. That way you'll have your neat little trend line without the outliers, but they are still visible on the graph.



Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?