Forecast |

Parameters |

ARTXP Model |

ARIMA Model |

Parameters |

ARIMA Model (.3 ADP) |

ARIMA Model (.8 ADP) |

ARIMA Model (0 ADP) |

This model appears to be predicting using the overall mean or something like it. It doesn't matter much because it doesn't seem that we can use this parameter to fix our issue. Let's go back.

Parameters |

There's another periodicity parameter called Periodicity Hint. This parameter allows you to tell the algorithm what kind of periodicity you think your data has. Our data has some definite peaks that appear in the middle and at the end of every year. So, why don't we set this value to {6, 12}.

ARIMA Model ({6, 12} PH) |

This model seems to account for the highs and lows much better than the previous ones. However, we have an issue that seems apparent now. Our data has a very recent upward trend. Our historic predictions don't seem to be accounting for this. The historic prediction model is built using two parameters, Historic Model Count and Historic Model Gap.

Parameters |

Historic Model Count dictates how many distinct historic models are built and Historic Model Gap dictates how many months each model will predict. For instance, the defaults were 1 and 10. This means that 1 model was built for the previous 10 months, using only data from more than 10 months ago. If we set these values to 2 and 5, we would get two models. The first model would see data more than 10 months ago and predict data until 5 months ago. The second model would see data more than 5 months ago and predict data until the final month. This would let us see how the model predicts for observations in the near future. Let's set these values to 3 and 2 and see what happens.

ARIMA Model ({6, 12} PH, 3 HMC, 2 HMG) |

As you can, these historic predictions are FAR better than what we had seen before. This leads us to believe that our model is pretty good, but only for a short period into the future. If we want to see further than that, we either need to use a different type of model (which would require a different tool) or get more data (which would require waiting).

This was a good investigation into the Microsoft Time Series Algorithm. Stay tuned for our next post where we will be talking about creating multiple models under a single structure. Thanks for reading. We hope you found this informative.

Brad Llewellyn

Director, Consumer Sciences

Consumer Orbit

llewellyn.wb@gmail.com

Director, Consumer Sciences

Consumer Orbit

llewellyn.wb@gmail.com

http://www.linkedin.com/in/bradllewellyn