Hierachical Time series
Locked
 
|
Dear Listers,
Is there any option in SPSS for us to perform hierachical time
series?
I know we can use hts package in R but anything that native to SPSS we can
use?
Thanks! |
|
|
Take a look at TCM (Analyze > Forecasting > Create Temporal Causal Models. It isn't quite the same as true hierarchical time series, but it might work for this. On Mon, May 2, 2016 at 9:00 AM, Chichi Shu <[hidden email]> wrote:
|
|
|
In reply to this post by Chichi Shu
Please give us a more detailed description of your statistical model, design, variables, etc. Ryan Sent from my iPhone ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
In reply to this post by Jon Peck
Thanks, Jon and Ryan.
Someone came to us with this really general question and I don’t have more
details. But given a typical hierachical modeling problem setting, I think HTS
might be the best choice.
Thanks,
Chi
From: [hidden email]
Sent: Monday, May 02, 2016 11:06 AM
To: [hidden email]
Subject: Re: Hierachical Time series Take a look at TCM (Analyze
> Forecasting > Create Temporal Causal Models. It isn't quite the
same as true hierarchical time series, but it might work for
this. On Mon, May 2, 2016 at 9:00 AM, Chichi Shu <[hidden email]> wrote:
|
|
|
In reply to this post by Chichi Shu
The short answer is that there is no “push-button” approach. SPSS Statistics gives you some good tools for automatic forecasting, as well as tools for aggregation, match-merging, etc. You would have to derive the aggregate series, develop the forecasts, and then distribute them “down” to the lower level series. Tony Babinec From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Chichi Shu Dear Listers, Is there any option in SPSS for us to perform hierachical time series? I know we can use hts package in R but anything that native to SPSS we can use? Thanks! ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
A few years ago, before TCM was born, I did a prototype hierarchical setup (using some real customer data). There were 2000 leaf units. The data was messy, because they didn't all start at the same time, and there was missing data in the middle in some cases. It was also a requirement that the model built for each leaf be available for inspection and the worst-performing models be easily found and summarized. It was also desired to be able to use some aggregate variables in the leaf-level modeling along with the leaf data. So the models were built and forecasts obtained using the usual Time Series Modeler procedure (TSMODEL). The issue then was how to aggregate/distribute the resulting forecasts. Aggregating the individual forecasts up just meant summing the leaf-level forecasts. If there had been a top-level model to be imposed on the leaf forecasts, it could have been done by scaling the forecasts in proportion so that they summed to the top level aggregates. In this case, the leaf models were all built independently. If you wanted to take account of other leaves in each leaf model, things get a lot more complicated, but hts has some capabilities for this. Also, it was necessary to worry about leaf models that failed because there was too little data for a particular leaf. Overall, this can be a messy business. On Mon, May 2, 2016 at 9:34 AM, Anthony Babinec <[hidden email]> wrote:
|
|
|
In reply to this post by Chichi Shu
I have engaged in interesting conversations in other forums where mixed models using other software seemed capable of re-creating some time series models while accounting for the hierarchical nature of the data. But some individuals view time series as a very specific type approach, understandably, which may not be appropriate for a mixed model. That's why I asked for details. Without details I would refrain from this conversation as it relates to mixed models. Ryan Sent from my iPhone ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
In reply to this post by Jon Peck
Jon, Your comments are well-taken. A few years ago, I encountered some data from a packaged good company. The time series were low-level “granular” with respect to brand, distribution point, packaging, time interval, and the like. The issue in a nutshell was on the relative merits of Forecast-First-Then-Aggregate versus Aggregate-First-Then-Forecast, or some hybrid. If you chose to develop forecasts first at the granular level, some series had very low volume, 0s or missing, and different start points. If you “naively” developed forecasts of the granular series and then aggregated them, the aggregated forecast series sometimes indicated implausible high volumes (client judgment) in a future time horizon. If you aggregated first on combinations of those grouping variables, the forecasts of the aggregated series indicated plausible future volumes and series behavior. Yes, the proportion-scaling you mention was used in this instance. Tony Babinec From: Jon Peck [mailto:[hidden email]] A few years ago, before TCM was born, I did a prototype hierarchical setup (using some real customer data). There were 2000 leaf units. The data was messy, because they didn't all start at the same time, and there was missing data in the middle in some cases. It was also a requirement that the model built for each leaf be available for inspection and the worst-performing models be easily found and summarized. It was also desired to be able to use some aggregate variables in the leaf-level modeling along with the leaf data. So the models were built and forecasts obtained using the usual Time Series Modeler procedure (TSMODEL). The issue then was how to aggregate/distribute the resulting forecasts. Aggregating the individual forecasts up just meant summing the leaf-level forecasts. If there had been a top-level model to be imposed on the leaf forecasts, it could have been done by scaling the forecasts in proportion so that they summed to the top level aggregates. In this case, the leaf models were all built independently. If you wanted to take account of other leaves in each leaf model, things get a lot more complicated, but hts has some capabilities for this. Also, it was necessary to worry about leaf models that failed because there was too little data for a particular leaf. Overall, this can be a messy business. |
| Free forum by Nabble | Edit this page |