The paper analyzes the possibility of using trend models to forecast the number of accidents and the number of fatalities and injuries in road accidents in Poland. In particular, linear and logarithmic models were analyzed, as well as the dependence of forecast effectiveness on the length of the time series. The impact of the forecast horizon on forecast quality was also analyzed.In the years 2005-2008, as part of research work [3, 4, 5], studies were conducted on the possibility of using trend models to forecast the number of accidents and the number of road accident victims in Poland. 11 classic trend models [6] were tested, initially considering time series of lengths 21, 9, and 8, which was determined by the number of available data and the “course” of the studied variables (maximum values occurred in 1997).
The analysis excluded the possibility of using time series of length 21, polynomial trends of degree 3 and 4, as well as hyperbolic I and log-hyperbolic trends. Classic ex-ante measures were initially used to evaluate the forecasts.
It turned out that the ex-ante evaluation of forecasts is significantly divergent from the ex-post evaluation; furthermore, some models with similar ex-ante evaluations produced significantly divergent forecasts. (Non-standard) ex-post evaluations were constructed and used to assess the quality (effectiveness) of the forecasts, taking into account the length of the time series – rolling evaluations. As a result of the analyses, it turned out that linear and logarithmic trends were characterized by the highest effectiveness. It should be emphasized that symptomatic forecasts, which include forecasts made using trend models, cannot “capture” sudden, unusual changes in the forecasted variable – which will be clearly demonstrated in this article, however, in many cases, they can be useful, providing surprisingly good forecasts many times. (…)
The article contains 18,939 characters.
Source: Logistyka Magazine