Support NeoGAF

[Stumpokapow]

I believe it is an aggregation of multiple bell curves, but for games as well as price points. You also have seasonal buying to factor in, as well. In short, it isn't one factor that drives purchasing trends but many. Going back to the Media-Create threads, you can see this clearly with Gunda/Golf/Yakuza/Smash/etc. bumps in addition to fluctuations with price. Monthly charts smear the game effects that are clearly seen in weekly charts. Cumulative charts also hide all but the most severe bumps.

To expand further; the purpose of using a regression is so that all of these bumps and sags, whether seasonal or number-of-weeks-per-month, or major game driven, or economic recession (apparently games are not currently impacted!) will be factored in equally.

How a model could better incorporate things is a different question.
- Exponential smoothing would better incorporate momentum so that old data does not have favour over recent trends. By contrast, it would mean that predictions are weaker at period where deductively we can say the trend is going to end.
- Dummy variables would be able to incorporate different month-lengths and holiday bumps, or any other bump that can be predicted with regularity.
- Curvilinear or exponential modeling would be able to better predict things, but the sort of data required to construct a curvilinear or exponential model that is nontrivially better than a linear model is just not available. You're talking about modeling sales at a "console lifespan level" and we just don't have enough datasets for that and the datasets we do have are tainted by all sorts of things like early deaths by choice, early deaths by financial requirements, early deaths due to competition, etc.

I do believe that the linear model is sufficiently good that it could only be substantially augmented by the use of mild exponential smoothing and the correction of the technical error in the regression in the article.

There are two big wildcards that can never be fixed:
- Supply constraints; except through channel checks, we can't measure supply constraint properly and so actually putting it in a descriptive-predictive model is pretty much impossible. If you add a variable for supply constraint, a human will still have to evaluate factors and add them in on a monthly basis.

- The mysterious impact of this month's software; again, a human has to convert "DEVIL MAY CRY 4" into a mathematical factor, and it has to be done on a monthly basis.

In terms of an ongoing machine-generated no-human-intervention model, linear regression is probably as good as it's worth getting. If anyone wants to make a curvilinear or exponential model, that's awesome too.

 

[IzumiK]

Wat? Sold less than SSBM.

Death of Nintendo Confirmed.

jk

 

[Stumpokapow]

Wat? Sold less than SSBM.

Death of Nintendo Confirmed.

jk

Unless it's a double-layered jk, you didn't read the graph right.

 

[XiaNaphryz]

Unless it's a double-layered jk, you didn't read the graph right.

Everything wrong with anything.

 

[jvm]

Whew. Meetings over, dishes done, kids in bed.

First, thanks for all the comments, as always.

Second, a comment about trying to improve based on the feedback here. I took last month's comment from Chris FOM (here) quite seriously and tried to focus just on the hard data. I had a very nice data point from NPD (thanks!) in the form of the Smash Bros. LTD data, so I didn't feel pressed to find something that people would really find eye-opening. I'd like to think that the result was a better article this month, but perhaps no one noticed but me.

Next, Stumpokapow has good points about how the linear modeling was done. The linear modeling is chosen for a couple of reasons: expediency (I have a deadline that must be fit into my real life constraints) and its reputation in modeling. In particular, if I start picking parameters for alternative modeling techniques (I have some ideas similar to what you've suggested, Stumpokapow, which I hope to have time to incorporate into my weekly column sometime this summer) then I'm in a position of having to justify them.

I feel that it would open me up to the same kind of criticisms leveled at ioi when it has been said that he weights various public data to account for what he feels are inaccuracies (real or imagined). Using a known model instead of something for which I have less rigorous justification is simply easier to defend. The heat surrounding sales is enough without throwing on extra fuel.

That said, there should be ways to estimate parameters without relying on ad hoc weights (like the 0.8, 0.2 you suggested). We'll see how much time I get in the summer to hash things out.

If I find a linear model for the last 12 months of data (April 2007 - March 2008) then here are the results:

The predictions for LTD sales as of December 2008 shift up for all consoles, but the conclusion that the Wii and Xbox 360 models cross around that time -- December 2008 -- is essentially the same.

As for my background, it is not statistics, but rather numerical analysis, and in particular numerical partial differential equations. The two fields overlap (lots of linear algebra, in particular) are not as similar as their names might suggest, in my opinion.

I'm sure there were other things to which I intended to respond. I'll leave it there for now and perhaps get back to other comments when I'm rested.

 

[dammitmattt]

Second, a comment about trying to improve based on the feedback here. I took last month's comment from Chris FOM (here) quite seriously and tried to focus just on the hard data. I had a very nice data point from NPD (thanks!) in the form of the Smash Bros. LTD data, so I didn't feel pressed to find something that people would really find eye-opening. I'd like to think that the result was a better article this month, but perhaps no one noticed but me.

I noticed that you used only hard data and I meant to comment on that earlier in the thread. I understand the criticisms that you open yourself up to when you try to extrapolate data (as well as the inevitable math errors), but I found it more interesting when you did this because you frequently brought up data points and discussion that was missed in the official NPD posts. Using just the data (minus the Smash bit), it's hard to find something that wasn't discussed to death in the NPD thread. I'm sure most of your readers won't care at all, but it does make it slightly less interesting for the NeoGAF crowd

 

[jvm]

Using just the data (minus the Smash bit), it's hard to find something that wasn't discussed to death in the NPD thread. I'm sure most of your readers won't care at all, but it does make it slightly less interesting for the NeoGAF crowd

Damn! I just can't win. :lol

Ah well. I'll keep plugging away at this job. Either I will get better...or I won't!

 

[Chris FOM]

I noticed that you extrapolated a lot less this month, and in my opinion at least the article was far stronger for it. The extrapolations were reasonable and based just on data we had, and you definitely had some nice extra tidbits (wow at Melee still doing 10,000 units) as well. I also think you use of linear extrapolation for the intersection line was totally fitting. The only wrench in the works is that the Wii's sales are supply rather than demand limited, so if Nintendo bumps the supply your own curves based on supply-limited sales will totally go out the window, but there's simply nothing that can be done about that. Until we see if March's numbers can be sustained (both from supply and demand) using them to recalculate the intersection between the Wii and the 360's respective installed bases is premature.

Great work this month, this is the best one you've done in a few months I'd say. It's at least the best since December's bare minimum.

 

[dammitmattt]

Damn! I just can't win. :lol

Ah well. I'll keep plugging away at this job. Either I will get better...or I won't!

Don't get me wrong...I still am a big fan! It's much more analysis than any of the other gaming websites perform. Keep them coming!!!

 

[nli10]

Don't get me wrong...I still am a big fan! It's much more analysis than any of the other gaming websites perform. Keep them coming!!!

Yup - keep them coming, I like the articles!