2026-05-10T08:07:51.625Z http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai

ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai:2:11414 2025-05-18T21:40:16Z 2

Gasoline and Crude Oil Price Prediction using Multi-headed Variational Neighbour Search-tuned Recurrent Neural Networks 2025 http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai/record/2/11414 https://link.springer.com/article/10.1007/s10614-025-10967-4 M. Kljajic V. Mizdrakovic L. Jovanovic N. Bacanin V. Simic D. Pamucar M. Zivkovic The rapidly growing worldwide demand for energy and finite reservoirs of fossil fuels have intensified interest in energy projection research. Artificial intelligence, particularly in time series forecasting, holds significant promise for enhancing predictions of both cost and demand, offering numerous prospective applications over various domains. Numerous variables, ranging from socio-economic conditions to distribution, supply, and international policies, exert influence on global price fluctuations. Thus, crafting precise forecasts necessitates consideration of these multifaceted factors. Through an examination of existing literature, a discernible gap emerges in the quest for advancement in this domain. Consequently, this study proposes to delve into the perspective of employing multi-headed long short-term memory (LSTM) models for gasoline and crude oil price prediction, an area largely untouched by multi-headed approaches. Moreover, recognizing the computational demands of such models, this research emphasizes the development of lightweight methodologies, characterized by a modest neuron within each layer and trained over a limited epoch count. Given the pivotal role of hyper-parameter selection in algorithm performance, an adapted version of the variable neighbour search algorithm is introduced to aid in tuning the model’s architecture and training parameters. A comprehensive side-by-side comparison is undertaken utilizing gasoline and oil data sourced from diverse public repositories, employing a variety of contemporary optimizers. The resultant outcomes are subjected to strict statistical scrutiny to ascertain the robustness and significance of the findings. article 10.1007/s10614-025-10967-4 1 36 1572-9974 Computational Economics