I’m often asked about how regional forecasts are constructed. The fact is that different outfits construct regional forecasts differently.
At the lowest level of rigor, a forecaster may just discuss broad trends, perhaps referring to historical data, but not provide a numerical forecast. These forecasts are not without value. Improved understanding the general sweep of things can help forecast users determine the major influences on their organization. Still, the information in this type of forecast is limited.
A slightly more rigorous forecaster may write down forecasted numbers, but the model generating the numbers is not formal, existing only in the forecasters mind. Again, there is information in this forecast. The forecaster could be astute and very well informed. The limitation is the lack of a structural model may result in inconsistencies, and explaining the forecast can be challenging.
The next level of rigor is a formal statistical model. This has the advantage of clearly identifying links and causality, which is valuable information.
There are a few outfits that will provide a statistical forecast for every county in the United States, but there are problems with these forecasts. One problem is that these models assume that links and causal relationships are identical everywhere. The better models will weight links differently, but they assume the same links everywhere. Also, because different data is available different counties and states, these models are limited by the data available for the most data-poor county.
A custom statistical model for a geography provides an improvement in information and rigor. However, even this type of model is not the best that can be done. These models can be constructed by analyzing data without ever visiting the region. However, visiting the region provides information that enables a forecaster to construct a better model. Regular visits provide opportunity to regularly improve the model.
Another, very serious, problem with purely statistical models is that these models are backward looking. I like to use Shasta County California to illustrate this problem. Suppose you have a statistical model of Shasta County California, but you don’t know that they found an owl in the woods, and the county’s lumber business will soon grind to a halt. You will have a bad forecast, because the purely statistical model can’t use this information. It only knows what has happened in the past, and even that with quite a lag for some types of data.
The most valuable type of model is a statistical model that has the ability to use information not yet reflected in the data. This is the type of modeling we do, through the use of what the industry calls add factors.
I’ll use Central Oregon as an example:
I visited Central Oregon for the first time about seven months before our first forecast. That visit was followed by two more visits. Another economist accompanied me on one of those follow-up visits. The purpose of these visits was to really understand Central Oregon’s economy, and that helped us develop a model that has performed pretty well.
Of course, building a model is not a one-time deal. We are constantly trying to improve our models. We also have to keep tabs on the economies we forecast. So, continuing the Oregon example, we make several trips a year to Central Oregon.
We do more than just visit a region a few times a year. We collect a lot of data for every region we forecast. Maintaining and cleaning these data is an ongoing process. We also follow a region through news and other media. For example, we subscribe to the Cascade Business News in Central Oregon.
Technology helps us too. We use Google to serve as a mail service. We receive e-mail links to every article or blog on the State, region, County, and cities on topics such as jobs and economy. We also receive e-mail links to articles and blog on the legislature. Going through those e-mails and links is part of our everyday scan for every region we forecast.
We also maintain relationships with citizens of the regions we forecast. These people will send us an e-mail or give us a call to tell us about events that they think could impact our forecast. Consequently, it is not unusual for us to know about an event before it is public.
We also have a State forecast for every region we forecast. You just can’t do a regional forecast without a State forecast. The State forecasts are performed quarterly, and like a regional forecast, these require constant monitoring. Sometimes, we have to do forecasts of other regions that may impact the region of interest. For example, a Ventura County forecast has to include a Los Angeles forecast, if it is to be accurate. Los Angeles County’s economy has a huge influence on Ventura County’s forecast. Similarly, a Central Oregon forecast requires a current California forecast.
There you have it, a glimpse into our sausage machine.