Neural titans in market prediction: MLP, transformer, & hybrid models across G-7 and China

Abstract

This study aims to conduct a comparative evaluation of eight state-of-the-art forecasting models – TimeMixer, PatchTST, iTransformer, NHITS, NBEATS, SOFTS, RMoK, and BiTCN – representing diverse deep learning architectures, neural basis expansion techniques, and hybrid approaches, for predicting stock market prices. We assess their performance across seven major global indices: the Shanghai Stock Exchange, S&P/TSX Composite, FTSE 100, DAX, CAC 40, S&P 500, and Nikkei 225, using rigorous metrics (MAE, SMAPE and RMSE). Our findings indicate that neural basis expansion models (NHITS, NBEATS) achieve superior overall accuracy (aggregated MAE: 0.013–0.014), particularly in North American and Asian markets. In contrast, transformer-based architectures exhibit market-specific strengths, with iTransformer delivering exceptional performance on Canada’s S&P/TSX (MAE: 0.003). Notably, European indices (DAX, CAC 40) present significant challenges, where BiTCN and RMoK underperform (MAE: 0.032–0.038), suggesting limitations in modelling abrupt volatility shifts characteristic of these markets. These results highlight critical regional performance variations and provide insights into architectural efficacy under diverse market conditions.