Twenty-Year Traffic Forecasts on 23,325 Ohio and New York Bridges: An Exact Growth-Gap Decomposition Shows Traffic Came in 26% Below the 2005 Forecasts

Every bridge record in the US National Bridge Inventory carries a forecast of average daily traffic for a future year, roughly twenty years ahead. I test the forecasts recorded in 2005 against the traffic counts in the 2025 inventory, for 23,325 Ohio and New York bridges that kept the same structure number and year built and had usable counts. Each forecast is interpolated geometrically to the year of the later count. Traffic came in below forecast on 77.4% of bridges. The median ratio of actual to forecast traffic is 0.744 and the ratio of total actual to total forecast traffic is 0.732; only 27.9% of bridges are within plus or minus 20%. The forecasts assumed a median growth of 1.49% a year, while realised growth had a median of minus 0.11% a year, and traffic fell on 52.6% of bridges. I prove that the log accuracy ratio equals the elapsed time multiplied by the gap between realised and forecast growth, exactly; the mean gap is minus 1.69 points a year over a mean of 18.2 years. Forecast growth explains little of realised growth: the correlation is 0.276 and the slope 0.43. The later count is less than proportional to the base count (elasticity 0.903), so forecasts that scale the base count over-predict most on the busiest bridges, a regression-to-the-mean effect that I derive under classical measurement error. The under-run is larger in New York (median ratio 0.677) than in Ohio (0.817), and survives excluding pandemic-year counts (0.710) and restricting to counts within the forecast horizon (0.763).