MS Time Series Algorithm -- Stationarity Transformation

Discussion in 'microsoft.public.sqlserver.datamining' started by paul, Apr 4, 2006.

  1. paul

    paul Guest

    Hello--

    How does the MS Time Series Algorithm check for stationarity? For non-stationary series, how does the algorithm transform the series to make it stationary?

    Thanks in advance,
    - Paul
     
    paul, Apr 4, 2006
    #1
    1. Advertisements

  2. Here's a response from Alexei Bocharov, our Time Series expert:

    As far as the existing ARTxp implementation used by the Microsoft Time
    Series algorithm in SQL 2005, it is addressing non-stationarity by
    performing splits on time when building the autoregressive tree. This is an
    addition to what is described in the "Autoregressive Tree Models for
    Time-Series Analysis" paper (
    http://research.microsoft.com/~dmax/publications/dmart-final.pdf).

    By allowing Time as a non-regressive input we try and isolate statistically
    different patterns on different tree branches.

    Here is an example from a retrospective "Economic Indicators" model that I
    have sitting around on my machine.

    Cross-predicting "New Motor Vehicle" production using other indices as
    predictors we might get

    [If] Date < 4/29/2004 11:26:15 AM
    New Motor Veh = 11.905 + 0.640 * New Durable(-1) + 1.073 * New
    Construction(-1) -0.560 * New Durable Other(-1) -0.624 * New Motor Veh(-1) +
    1.022 * New MV Bodies(-1)

    [If] Date >= 4/29/2004 11:26:15 AM and New Motor Veh(-1) >= 39126
    New Motor Veh = 28710.713 + 1.243 * New Motor Veh(-1) -0.350 * New
    Durable(-1) -0.277 * New Construction(-1) -1.582 * New MV Bodies(-1) + 0.453
    * New Durable Other(-1)

    This shows ARTxp model detecting different regressive patterns in "New Motor
    Vehicle" production 4/29/2004

    --
    -Raman Iyer
    SQL Server Data Mining
    [This posting is provided "AS IS" with no warranties, and confers no
    rights.]

    PLEASE POST NEW QUESTIONS AT
    http://forums.microsoft.com/msdn/showforum.aspx?forumid=81&siteid=1
     
    Raman Iyer [MS], Apr 7, 2006
    #2
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.