MODEL FORECASTING HARGA KOMODITAS UDANG DI PT. X DENGAN MENGGUNAKAN PEMBELAJARAN MESIN (MACHINE LEARNING)

• Main Author: Ridho Zul Khoiri, Muhamad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/80834
• Institution: Institut Teknologi Bandung

This study aims to build a prediction model of shrimp prices one hundred days ahead for PT. X, which faces challenges in setting accurate forward contract prices for vaname shrimp commodities due to high market volatility. The predictor variables used are derived from historical transactions of vaname shrimp from PT X for 499 days (9,182 transactions) and external variables, such as macroeconomic factors and shrimp import-export prices. Data processing includes cleaning, transformation, and division of data into training and testing data. The models built are multiple linear regression (MLR) with lasso regression techniques, random forest regression (RFR), extreme gradient boosting (XGBoost), and aggregation of the three models with the averaging method. The construction of the MLR lasso model began with selecting relevant independent variables and overfitting with an optimal alpha value of 257.721. The model showed a relationship between shrimp prices and internal factors (shrimp size, date), external factors (global market conditions), and regional factors (location of South Sulawesi and Southeast Sulawesi), with an RMSE of Rp5,168/kg in the training data. The RFR and XGBoost models were evaluated through experiments that considered hyperparameters (n_estimators, max_depth, min_samples_split, min_samples_leaf, gamma, colsample_bytree, lambda, alpha, and learning_rate). Using the Bayes search algorithm, the best RMSE obtained for the RFR model was IDR3,259/kg with the hyperparameter values in order (207, 10, 33, 35), while for the XGBoost model it was IDR4,578/kg with the hyperparameter values in order (89, 3, 0.057, 0.0925, 600.815, 389.357, 0.0296). The MLR lasso, RFR, and XGBoost models provided more accurate predictions than the company's in-house model with the RFR and XGboost models showing the best performance (5.29% and 5.02% respectively) on the testing data. Paired t sample tests proved the use of more complex models and the addition of external variables improved the prediction accuracy of the models. A significant improvement in accuracy was seen from the regular MLR model to MLR lasso. The addition of external data was also significant in improving the prediction accuracy of the model, although the effect was marginal (MAPE 5.84% to 5.81%). Backtesting shows that the prediction model is more selective in initiating forward contracts by selecting 373 contract positions (out of a total of 976 historical decisions) for the XGB model, yet has the potential to increase win-rate (from 37.19% to 75.87% with the adoption of the RFR model), Sharpe Ratio (from -0.26 to 0.33 with the adoption of the XGB model), and potential profit (from -Rp6.86 billion to Rp2.18 billion with the adoption of the XGB model). The model can also reduce the risk of loss per transaction (from IDR 16.7 million to IDR 4.9 million), signalling potential improvements in performance and risk management. With the adoption of the more accurate prediction model generated by this research, PT X can perform better forward contract pricing, minimize risk, and proactively adjust the company's strategy according to future market dynamics.