Crude Palm Oil Price Prediction Using Simulated Annealing-based Support Vector Regression (SA-SVR)

Abstract

Palm oil is one of the major export products of Malaysia. Predicting the price of crude palm oil futures (FCPO) traded on BURSA Malaysia Derivative is essential as agricultural markets have an inherent tendency towards instability, and thus are more vulnerable to price shocks than other industrial sectors. Hence, if the price of the futures contract on crude palm oil can be forecasted accurately, many parties such as farmers, refiners and distributors can manage the risk of price fluctuations through FCPO. This study proposes the metaheuristic and machine learning hybridised model of simulated annealing-based support vector regression (SA-SVR). The SVR in this model produces close price predictions of the FCPO with minimum deviation from the actual value with the help of SA, which first determines the best hyperparameter set to be utilised in the SVR. Although the proposed Radial Basis Function (RBF) kernelised SA-SVR model inputs only 10% of training data due to memory overload issues, it has produced a satisfying prediction result with an average execution time of 2 minutes and 34 seconds. The model performance was analysed further by using different ratios in data splitting, varying temperature combinations for the SA algorithm and initiating the parameter search based on the previous best hyperparameter set. Results show that keeping the test size constant and extracting more historical data on FCPO price for model training is better than varying train-test split ratios. The temperature schedule strategy showed that different initial and minimum SA temperature combinations affects the overall optimisation results. The best combination was the initial temperature of 100 and minimum of 40. In addition, the number of temperature reductions and average execution time to reach the best state decreases when the starting point of the parameter search space is close to the best values.