High-Density Polyethylene Film Price Forecast in Southeast Asia Market with Deep Learning

Abstract

Neural networks (NN) have been used for over a decade to predict time series data, with various algorithms including linear and non-linear models. Forecasting and assessing polymer market prices are crucial for plastic resin producers due to the complexity and uncertainty of resource availability. This encompasses feedstock planning, raw material procurement, technological advancements for product transitions, sales planning, pricing strategies for commercialization, and investments driven by macroeconomic factors. Previous literature primarily utilized numerical data as input for deep learning models. This research contended that structured data by itself was inadequate for models to precisely predict outcomes in the volatile, uncertain, complex, and ambiguous (VUCA) environment.  Three deep learning architectures, Long-Short Term Memory (LSTM), Encoder-Decoder, Temporal Convolutional Network and Recurrent Neuron Network (TCNRNN), were reviewed in this research to determine the most effective architecture for analysing structured data. Additionally, Natural Language Processing (NLP) was implemented in this research to gather market sentiment and enhance forecast accuracy. The study utilizes commodity market price announcements, economic indicators, and insight reports from reputable publishers. The study utilizes commodity market prices, economic indicators, and insightful reports. All information was obtained from a reputable publisher. The results were compared with the legacy model, which involved a human analyst and a linear regression model. Model performance was assessed using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Symmetric Mean Absolute Percentage Error (SMAPE), and ANOVA. The linear regression forecast together with the human analyst model has an acceptable accuracy with a MAPE of 45.1%. Neural networks containing sentiment analyzers have been found to surpass the performance of human analysts and a linear regression model, with a MAPE of 17.1%.