Epsilon Archive for Student Projects

Abstract

At a time when data are an integral part of many industries and the world as a whole, driven by a
digital dominant era, there is an emerging discussion surrounding the level of value captured. Using
data accumulated in the forest management system and the surrounding databases of the forestry
semi state Coillte, my aim was to research if there was any type of data that could potentially be
used to identify and understand inaccuracy in long- and short-term forecasting. The Company’s
Tactical and Strategic forecast volumes for 2018 were used in conjunction with the actual harvest
volume from weigh-bridge measurements and roadside stocks in order to understand the current
extent of over- and under-estimation. To achieve this, the methods of linear and stepwise backwards
logistic regressions were used. The linear regressions based on the percentage of difference of the
forecast volumes towards the actual harvested volumes were inconclusive. The logistic regressions
were produced using eight binary response variables based on over-and under-estimation. For each
forecast type they included, a dataset with all species and product volumes, a dataset with only the
dominant species volume (Sitka Spruce/Picea Sitkensis), and datasets using the most valuable
product (large sawlog) with total volumes and volumes of the primary species only. The predictors
consisted of landscape and geographic variables; namely elevation, slope, aspect, country segment,
distance from coast, latitude, soil type and roughness. The results showed that over-estimation is the
most common form of forecast bias with the tactical forecast models being the most accurate using
the predictor variables, Elevation, Roughness and Soil Type. The variables, Aspect, Segment within
country and Slope were shown to be the least valuable for prediction. Thus, these aspects should be
taken into account when researching forecast bias at the planning level.