How is Powerset predicting growth?
Over the past month, Powerset has been working on building a modeling framework to predict our own growth once we launch. Of coarse, modeling is a tricky business and the trick is to build a model that can accurately predict reality. Many models that I have seen (and yes I Steve Newcomb am an excel modeling junky) either take a top down approach or a bottoms up approach. The problem with a top down approach is that it often misses things like the SlashDot effect or the BlogStorm effect. However, the problem with a bottom up approach is that you often end up with endless assumptions with many assumptions having a high sensitivity to the overall model - ultimately ending up with a model that is extremely complex and not a true picture of a real scenario.
At Powerset, we have created a modeling system that comprises both a top down approach and a bottom's up approach. The way we did this was by putting top down constraints on our model that limits the outerbounds of any calculation to be within the norms of similar competitors growth curves. I.E. we can constrain our system to be +/- a certain boundry percentage of any competitor based on their real growth curves. For the bottom up approach, we built a model that computes population attrition and organic growth. We track each population over time as its own unique and distinct population using typical attrition and unique search growth algorythms.
To test our model, we built a model predicting the growth curve of unique searches for Snap.com after its release of its new Snap Preview Anywhere release and its re-launch of its product. The graph above represents the results. You can see Snap's real unique users (shown with the blue line - take from Quantcast) and our prediction (shown in the blue background.)
While I would say this doesn't mean that we have a crystal ball, it does mean that we have a pretty cool way of modeling growth that can help us plan for how many machines we will need to buy. It's a little bit of a peek inside of Powerset to see that we are total geks when it comes to anything we do.
I don't know if anyone would find this interesting but I thought I would post it anyway.
Comments
Thanks for being transparent on this topic. Models based on reality will always be better estimates than those based on abstract theories.
And yep, there are people out there who are very interested in this sort of thing.
(I do capacity planning at a familiar photo sharing website, so these issues are a large part of my job)
Posted by: john allspaw | May 5, 2007 09:50 PM
It's actually quite interesting.
The big question is how long was your training set and how far did you try to predict.
Posted by: Yaniv | May 6, 2007 12:49 AM
Modelling Snap as your bottom up model? Interesting given that snaps growth was driven by viral widget distribution. Presumaby Powerlabs will tie in your viral component - or do you have another plan? :)
Posted by: Paul Reilly | June 15, 2007 03:59 AM