Slimebeast said:
So can someone explain ensemble forecasting and how it applies to these professional video game sales prediction graphs, and the difference between installed base and unit sales, instead of being sarcastic.
response:
Although not quite correctly, I would describe ensemble forecasting as simply the use of a weighted average of forecasts. Now assume these forecasts have different variances and averages at each point in time and are skewed in some direction such that the median does not equal the mean (the risks are on the up/downside). You want to find the set of weight functions which reflect you confidence in a various models under various circumstances.
I don't know that this is being used here for video game sales prediction. But ensemble forecasting is in vogue for meteorology and market prediction and there is enough money involved for people to hopefully take things seriously enough to hire experts and it would explain the graph.
If you would like a more detailed explanation I would be happy to point you to the foundational literature or elaborate on anything specifically you are unclear about. Your level of mathematics training would also help me in any clarification.
Bodesatva said:
However, your insistence that this is an HMM is a circuitous -- and not entirely accurate -- claim. Obviously all market trends are, in a way, like HMMs; we know what's selling, but we're never entirely sure why (this makes our example fundamentally different than an HMM, by definition, but let's relax the definition slightly since it's not too far off) We have possible reasons, however.
response:
You I must say have been a consistently insightful poster. And I am not just saying that to be nice. I am not that nice of a person.
You caught me. Kings to you. When I was initially typing that I new I wasn't being very clear -- or less generously promulgating outright false-hoods -- in the name of expediency. I admit a combination of arrogance and laziness conspired against what needed to be done.
You are correct, in that an HMM assumes you know all possible states of your hidden model and all possible variables to be completely accurate. Relaxations of this assumption exist, but I am not a big believer in them as the whole point of HMM is it's simplicity. The unrealistic assumptions of course don't stop there, for example it assumes your hidden variables have the Markov property. The point I was trying to make was the following:
1) Assume as is claimed we know accurately what all observable parameters are (I don't dispute that this is mostly true, I personally use vgchartz for investing).
2) Lets also assume we know what all the underlying variables are (ie. X1(Y) is the median utility derived from a console where Y=sup(the sequence of the "party" factor of all games on a system). Ok so technically we have a Hierarchical HMM, but it's not that different.
3) Hidden variables have the Markov property.
4) We know the parameters of hidden variables in the future.
Even with all four of these -- a pretty generous set of assumptions -- unless someone on this forum is running some variation of the foward-backward algorithm they have no idea what the relative probabilities of sequences of future observations will be.
The distinction should be made between this and Kant's "unknown unknowns" which however can be dealt with in an ensemble of HMM-like models with good old fashion PCA which only assumes the underlying "truth" of the universe does not change (if you have infinite samples).
Bodesatva said:
Again, it is always true that we know what's selling, but can never precisely know the reasons why.
response:
Well the correct way of doing modeling this market would be using an ensemble method imho. The models which rely on prior knowledge would be combined with models derived from classifiers trained from past data (and using cross-validation to get a handle on what your error rate is) essentially allowing you to have a reasonable fall back when you leave the domain for which your "laws" are known to function reasonably. It's "just" a matter of optimizing -- in the L-infinity norm, but hey if it was easy everyone would do it -- your weights in your model. As you probably know neural networks (a scam imho), SVMs, and really just the plain old SVD (known that statisticians as PCA) all find optimal approximations of the underlying model without knowing anything beyond inputs and outputs.
Bodesatva said:
This is particularly true when a product in a well defined industry disrupts the understood values of the market (in the case of video games, those understood values would be multimedia features, graphics, and online functionality). In such cases, I'm sure you're aware that concrete, empirical data is particularly and emphatically relied upon, and that is the condition the video game industry is currently in.
response:
I think these changes could easily have had more of an impact than any features. I don't think it is over the top to suggest that Japan's shrinking and aging population has effected the video game market there. Currency movements caused by a number of things such as faster relative growth and preemptive moves to wean EU states off of oil which has led to muted inflationary effects of recent oil spikes. Probably has had significant impact on the relative pricing and thus sales in EU countries. How about demographic shifts in percentages of english speakers. Notice how Microsoft (with not the greatest record when it comes to localization) seems to do better in english speaking countries and countries which in my experience have high levels of enliglish speakers. Sony seems to have always concentrated heavily on localization (an overlooked advantage of BR, but only after they did away with region locking). Look at Sony performance in the UK > Neatherlands > Italy. This seems at least as plausible as any sort of vague "brand" value theories.
Bodesatva said:
Therefore, while it is unfortunate that the site offers little insight into demographic shifts or economic growth/depression, those factors are of significantly less importance in a disruptive environment.
response:
warning non-rigorous rant-- you may want to skip
No? What if recession is causing many to opt out of the $2k gamming rig and pushing the cheap consoles. Quite frankly I don't see the market as disruptive at all. I see each company having their own sales cycles all of which are weakly interacting. Adjusted for inflation all of the consoles are dirt cheap when compared to 1985 and are even cheaper as a percentage of disposable income. Hence demographics seems to allow for the peaceful coexistence of 3 consoles.
However something which could be considered disruptive is the slowing of advances in computer engineering. It happens to every field, we are running out of low hanging fruit (just look at pure math or physics for a glimpse of the future). That in my opinion is the real change this generation. If the generation last significantly longer, and more people can afford multiple systems, then who is to say that all three consoles wont continually switch off as "leader".
-- end rant --
I want to be clear that in no way am I attacking empiricism (at the moment :P). Nor again am I attacking this site, it does what it does well. I want to emphasis that. I have trolled this site for some time now and been constant witness to the truth of the saying "a little bit of knowledge in the wrong hangs can be a dangerous thing". At first sight of the two words "cherry picking" a sense of despair descends upon me. It's all very clock-work orange-ish.
Bodesatva said:
But again, it's certainly not insignificant, and it's important that you bring it up. If you have access to any of the aforementioned data, please do provide it! It would certainly help, and you do seem to be reasonably aware of the variables considered in operational forecasting.
response:
I have no time to collect any of the data as it's publish or perish and I have a thesis to finish. I have a degree in economics and am currently pursing a phd in Computational Science specializing in HPC particularly with clustering algorithms. I do have experience working with forecasting models (fault prediction at IBM) and have dealt with some of the meteorological models people develop here. If someone wanted to go through the dirty work of collecting all the data from various sites and formating them in a nice way, I could dedicate some time to developing a decent forecasting model we are always looking for good applications of our fancy algorithms here.. We could probably get some decent publications out of it and it might even be worth something. Hell, you could probably turn it around and use videogame sales as a predictor for all matter of other things as well. Videogame sales as a predictor for interest rates, hmmm :P
PS: I apologize for the wall of text
