Building a predicting model for a time forcast

Question

Hello Everyone,

Would you please guide me through my analysis:

I have an outcome for the years 2000-2019 which is based on multiple predictors (a,b,c,d,e)

1- I want to create a model in which I predict the outcome in 2020 where I have all the data for 2000-20019 plus I have the values for (a,b,c,d,e) in the year 2020, and I need to predict the outcome in the mentioned year and then compare it to the observed value.

2- How about predicting the outcome in the year 2020 without having (a,b,c,d,e) for the year 2020? Is it possible and valid if I built up a robust model???

Note: all my variables are continuous variables

peng_liu · Answer

There is a class of model under the name Transfer Function Model

This class of models expect that you have a series of Y (your outcome at all time points), and series of inputs (your a, b, c, d, e measure at all time points).

E.g. your have y[0], y[1], ..., y[n], y[n+1], and a[0], a[1], ... a[n], a[n+1], etc.

This class of models can help you to predict Y at a future time t, using past values of Y and past-AND/OR-current values of your inputs.

E.g. predict y[n+1], using y[0], y[1], ..., y[n], and a[0], a[1], ... a[n], a[n+1], and so on.

If you have a[n+1], at a future time, great, things just work out.

If you don't have a[n+1], a technique is to put your predicted value of a[n+1] in the place. To have your predicted value for a[n+1], you need a separate model on the series a, to get a[n+1].

The model is under a modeling framework known as ARIMA (autoregressive integrated moving average). But the steps can be applied to other general modeling approaches if the normality assumption is terribly violated, as long as you can formulate the prediction of future values using the past ones..

statman · Answer

First, welcome to the community.  Is this a real situation or hypothetical?  How was the data for those years acquired? (e.g., observational data, designed experimentation, directed sampling).  How have you evaluated measurement errors?

If it is real, how did you identify a-e as the predictor variables for the years 2000-2019?  Do you have hypotheses that support the relationships you have quantified? How well does the model you are using work for any of those years? Have you looked at the residuals for any years based on your model?

The comparison in question 1 is residuals analysis.

I don't understand question 2?  You want to predict the results of 2020 without a model?  This is non-sensical?  Are you looking to create a neural network to predict the data without knowing a model?  What do you mean by a robust model?  Typically it means the model is resistant to or less impacted by changes in variables outside of the model (noise).  Or the model effects are similar over changing noise (extraneous variable changes) or there are no noise-by-factor interactions.

P_Bartell · Answer

There are any number of modeling methods that might help you solve the problem. A good place to start is the Fit Model platform. Standard least squares personality. I suggest taking a look at the online documentation associated with this platform for specific details. Here is a link: Fitting Linear Models

Note there are many other platforms that might be more applicable wrt to solving the problem at hand.

All JMP modeling platforms have an option in the platform report to save the prediction formula to a data table as a new column. So once you find your 'robust model'...whatever that means in practice...just save the prediction formula to the data table and you'll get the 2020 prediction in that row. Or you could manually get there as well using the report prediction profiler. All modeling platforms have some form of Profiler for precisely doing what you want.