| Summary: | Salmon farming is the largest growing food supply sector in the world. Alongside the growth, the industry is becoming more competitive and is strengthening its position in the capital markets. However, the salmon price is characterized by high volatility, which imposes uncertainty and additional costs on the value chain of salmon production. The participants of the salmon market can benefit from a reliable price model, as it can be used to improve decision making regarding the operational and financial aspects of the industry that are subject to the price uncertainty. This includes the timing of salmon harvest, required machine capacity, investment decisions and stock valuation. This study aims to provide a framework for predicting the spot price of the Atlantic salmon, represented by the NQSALMON index. I present a new model, called the ARDL-PLS model, which is used to predict the 3-,6- and 9-month ahead spot price. The ARDL-PLS model is a combination of the autoregressive distributed lag model and the partial least square regression. Each of the three months to be predicted are appointed a designated submodel, which are independent on the other submodels. The covariates used in each submodel are selected from a database of 61 candidate predictors, using a genetic algorithm (GA) for feature selection. The GA-search is constrained to only select subsets of predictors that have a unique cointegrated relationship with the salmon price, which enables the use of non-stationary data in the ARDL-PLS model. The out-of-sample results of the ARDL-PLS model is compared to the results of an ordinary least square regression (OLR) model, which is implemented on stationary transformed data There are several encouraging results of this study. The genetic algorithm functions well as a feature selection tool, as it is relatively quickly able to select subsets of predictors that cointegrate with the salmon price, and that results in a favourable goodness-of-fit. Generally the ARDL-PLS model is able to explain a large degree of the variance in the salmon price, and the predictive accuracy of the ARDL-PLS model is better than the OLR model on all forecast horizons. The two models perform relatively similar in the 3 month-ahead prediction, but the ARDL-PLS model excel in the longer horizons. I attribute the performance difference between the OLR model and the ARDL-PLS model with two factor: (1) In contrast to the ARDL-PLS model, the OLR model requires the data to be stationary transformed, which removes the long term information in the data. (2) Unlike the OLR model, the use of intercorrelated predictors in the ARDL-PLS model does not affect the stability of the estimated regression coefficients, as the ARDL-PLS model transforms the covariates into orthogonal uncorrelated factors. Finally, the results indicate that the use of exogenous variables in regression based salmon price models, significantly increases the prediction accuracy.
|
|---|