Oops, they did it again.
As we mentioned earlier in the week, way back in 1991 the Washington State Department of Transportation recommended that engineers forecast future traffic volumes by finding a recent period of steady traffic growth, and simply assuming that this growth would continue. Although that approach probably worked well enough back in 1991, it’s been failing the region badly for the last decade. Surely by now traffic engineers know better!
Or, perhaps not. Take, for example, the Oregon Department of Transportation (ODOT) “Analysis Procedures Manual,” first published in 2006 and most recently updated in 2011. It calls for almost the exact same procedure as was found in WSDOT’s 1991 manual: running linear regressions from past trends to forecast future traffic growth. For details, see Section 4.6, or just read this excerpt:
The historical trends method uses traffic volumes from previous years to project future volumes. This method assumes that the future growth trend will be similar to the historical trend. It is used mainly in rural or small urban areas where significant growth is not anticipated.
[UPDATE: I had a nice call from someone at ODOT, who thought that our characterization of their methods was unfair. ODOT only uses the historical trends method in limited circumstances, when more sophisticated methods simply aren't available. Typically---and especially in urban areas---ODOT relies on other forecasting methods that are detailed in their manual. I think that's a fair critique: the traffic manual really does call for using a simple regression on past trends to predict future traffic volumes, but it also details other methods for urban areas or in places where significant growth is anticipated.]
Two things make this approach particularly curious. First, traffic volumes on Oregon state highways have flat-lined for about a decade. Yet the traffic manual suggests dipping into data from the 1990s or earlier—a time when gas prices, income growth, and basic demographics were strikingly different from today’s reality.
[ANOTHER UPDATE: The gentleman from ODOT reminds me, correctly, that dipping into past data doesn't automatically mean projecting growth. If traffic has been shrinking, ODOT's methods might suggest that traffic will continue to shrink! Yet I played with the method myself, and it's pretty clear that a decade of steady growth (fairly common in the 1990s) followed by a decade of no growth (fairly common in the 2000s) can produce a linear projection for steady future growth that meet's ODOT's stated statistical standards for forecasting. To me, this all suggests that using linear regressions to forecast past trends is inherently problematic: it may be the best method available in some circumstances, but that doesn't mean it's any good.]
Second, the manual recommends forecasting rising levels of traffic even for “rural or small urban areas where significant growth is not anticipated.” In other words: even in places where employment and population growth are at a standstill, it’s still perfectly acceptable to assume that traffic volumes will grow.
Using linear regression as a forecasting tool has a fatal flaw: it assumes that the future is going to look exactly like the past. As we’ve documented time and again, automobile traffic is no longer growing like it used to. Yet all too often, traffic forecasters are still relying on a 20-year-old mindset to make guesses about what’s coming next.