| dc.description.abstract | Industrial salmon farming is becoming an increasingly important industry, both globally and
in Norway. One of the main risk factors in salmon production is the highly volatile spot price,
so access to high-quality price forecasts could prove immensely valuable throughout the value
chain. In this thesis, we therefore attempt to make accurate and reliable forecasts of the salmon
price 12 months ahead and assess the potential economic value of such forecasts. We chose to
use tree-based models for this task, and the models applied were decision trees, random forests,
and xgBoost, with- and without seasonal adjustment.
The tree-based models displayed different levels of forecast accuracy, however all models
performed better than the seasonal naïve benchmark. Measured by mean absolute error and
mean squared error the best performing model was random forest, followed by xgBoost and
then decision tree. Overall, the seasonally adjusted random forest performed best, with a
directional accuracy of 82%, implying that the model correctly predicted up- or down price
movements around 8 out of 10 times. We found that the potential economic value of such
forecasts to SalMar, the third largest salmon producer in Norway in 2021 with a market share
of around 11%, could be 51.2 million NOK in additional earnings, corresponding to a 2%
increase compared to their total 2021 earnings. | en_US |
