EVE Search - Plex Prices

Aoife Fraoch wrote:

Teckos Pech wrote:

Aoife Fraoch wrote:

Teckos Pech wrote:

There is little to no seasonality in the PLEX price time series. I have downloaded the data and checked for both monthly and weekly seasonality and found none.

As for bubbles and the trend in PLEX prices....

As for bubbles, here is the thing about them...they are very, very hard to predict.

Lets suppose that some guy named Bob had found a way to predict perfectly when a bubble was about to burst. Bob would then sell short at the peak and then buy when prices dropped dramatically making a killing. How often could Bob do this? Maybe 2-4 times then everyone would know...watch Bob. As soon as Bob starts selling short....oh ****. Get out!!! Bob, to try and maintain his profit margin might sell a bit earlier...but then so would everyone else. And this logic would repeat itself all the way back to the point that the bubble started to form.

Predicting bubbles and their peak is not possible. Of course some people appear to have gotten it "right" by consistently going against the consensus, but so what? Doing that means that eventually you'd be right, but the vast majority of the time you'll be wrong. People who then think such prognosticators are some sort of genius simply help assure why bubbles are so hard to predict.

Is the current PLEX price a bubble? I don't know. Maybe. I doubt it, but then again nobody really knows.

There does seem to be yearly seasonality and a general upwards trend. Thanks for posting your data as well, I had a bit of a look at it. I have been meaning to do more with timeseries and it was fun playing with the forecast library.

http://rpubs.com/randomGraphs/117046

How many models did you fit? That just seems odd. One year high the next low. I could buy something more like a traditional seasonal patter what with school/summer etc. But a year over year fluctuation....?

And did you try including some sort of autoregressive error term? Looking at the remainder graph and your Durbin-Watson statistic (although I don't have access to my texts with the tables for the Durbin-Watson tables) might want to consider adding in an autoregressive error term.

Sorry for the delay, irl has been a little unfriendly of late. Nothing serious, just time and energy consuming.

Frankly this was just something I threw together because it looked interesting, and it did seem that when you have daily data, with some of the trending that is bound to happen around games, you would get some weird cycles turning up and there would not be just one pattern.

Personally I am templed to say that the last year has been weird and has not behaved like the previous ones. And even just mentioning being able to buy sp with plex will continue to make it odd.

The first time serious decomposition I did was using the stl function (using loess) with the seasonal window of "periodic" it takes the mean instead of smoothing per a supplied lag. The function can take a number of other parametres like a trend window and so on. I used this with a number of stated seasonal periods as well (7 and 365.25), but thats in the code.

The tbats function from the forecast library is an interesting one. For one thing it has an insanely long name (TBATS model (Exponential smoothing state space model with Box-Cox transformation, ARMA errors, Trend and Seasonal components)).

The options I had used were seasonal periods for both 7 and 365.25 on daily data with box.cox to true. If this was null then it would have used selected if they were to be used based on AIC.

trend and damped trend were not selected explicitly though damped was used. I should have included this in the original output:

> PlexAverage.tbats$parameters
$vect
[1] 1.246931e-06 8.484456e-01 8.181487e-01 -4.206490e-02 -6.835356e-07
[6] -3.498706e-04 -4.151820e-04 6.235535e-05

$control
$control$use.beta
[1] TRUE

$control$use.box.cox
[1] TRUE

$control$use.damping
[1] TRUE

$control$length.gamma
[1] 4

$control$p
[1] 0

$control$q
[1] 0

arama.errors was set true and used and fitted by AIC.

hope this answers your questions, its worth just looking at it yourself. I am relatively new to working with time series data anyway.