Take the 2-minute tour ×
Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It's 100% free, no registration required.

I need to forecast the following 4 variables for the 29th unit of time. I have roughly 2 years worth of historical data, where 1 and 14 and 27 are all the same period (or time of year). In the end, I am doing a Oaxaca-Blinder style decomposition on $W$, $wd$, $wc$, and $p$.

time    W               wd              wc               p
1       4.920725        4.684342        4.065288        .5962985
2       4.956172        4.73998         4.092179        .6151785
3       4.85532         4.725982        4.002519        .6028712
4       4.754887        4.674568        3.988028        .5943888
5       4.862039        4.758899        4.045568        .5925704
6       5.039032        4.791101        4.071131        .590314
7       4.612594        4.656253        4.136271        .529247
8       4.722339        4.631588        3.994956        .5801989
9       4.679251        4.647347        3.954906        .5832723
10      4.736177        4.679152        3.974465        .5843731
11      4.738954        4.759482        4.037036        .5868722
12      4.571325        4.707446        4.110281        .556147
13      4.883891        4.750031        4.168203        .602057
14      4.652408        4.703114        4.042872        .6059471
15      4.677363        4.744875        4.232081        .5672519
16      4.695732        4.614248        3.998735        .5838578
17      4.633575        4.6025          3.943488        .5914644
18      4.61025         4.67733         4.066427        .548952
19      4.678374        4.741046        4.060458        .5416393
20      4.48309         4.609238        4.000201        .5372143
21      4.477549        4.583907        3.94821         .5515663
22      4.555191        4.627404        3.93675         .5542806
23      4.508585        4.595927        3.881685        .5572687
24      4.467037        4.619762        3.909551        .5645944
25      4.326283        4.544351        3.877583        .5738906
26      4.672741        4.599463        3.953772        .5769604
27      4.53551         4.506167        3.808779        .5831352
28      4.528004        4.622972        3.90481         .5968299

I believe that $W$ can be approximated by $p\cdot wd + (1 - p)\cdot wc$ plus measurement error, but you can see that $W$ always considerably exceeds that quantity because of waste, approximation error, or theft.

Here are my 2 questions.

  1. My first thought was to try vector autoregression on these variables with 1 lag and an exogenous time and period variable, but that seems like a bad idea given how little data I have. Are there any time-series methods that (1) perform better in the face of "micro-numerosity" and (2) would be able to exploit the link between the variables?

  2. On the other hand, the moduli of the eigenvalues for the VAR are all less than 1, so I don't think I need to worry about non-stationarity (though the Dickey-Fuller test suggest otherwise). The predictions seem mostly in line with projections from a flexible univariate model with a time trend, except for $W$ and $p$, which are lower. The coefficients on the lags seem mostly reasonable, though they are insignificant for the most part. The linear trend coefficient is significant, as are some of the period dummies. Still, are there any theoretical reasons to prefer this simpler approach over the VAR model?

Full disclosure: I asked a similar question on Statalist with no response.

share|improve this question
    
Hi there, could you give some more context around the decomposition you wish to do, as I have not seen it applied to time series data? –  Michelle Mar 15 '12 at 6:38
    
I am breaking up the change into components in the following way: $W^{′}-W=p^{′}∗(w^{′}_{D}-w_{D})+(1-p^{′})∗(w^{′}_{C}-w_{C})+(w_{D}-w_{C})∗(p^{′‌​}-p)+(\epsilon^{′}-\epsilon)$, where primes denote current value of the variables. –  Dimitriy V. Masterov Mar 15 '12 at 13:09
    
hmmm, how about exclude the outliers first, before regression? –  athos Aug 31 '13 at 14:37

1 Answer 1

A model for this data isenter image description here

1: http://i.stack.imgur.com/LVn5D.jpg with supporting specifics enter image description here and enter image description here . The residuals appear to be random enter image description here with an original acf of enter image description here and a residual acf of enter image description here

share|improve this answer
2  
Using 10 parameters on 28 values to forecast the 29th would seem, on general principles, to be a gross overfitting. What steps have you taken to correct for that possibility? And how, specifically, does this model address the two sets of questions that have been posed? –  whuber Nov 29 '13 at 17:50

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.