LSTM-BASED PREDICTIVE MODELLING FOR GREEN AND NON-GREEN CRYPTOCURRENCY PORTFOLIO OPTIMISATION: A CROSS-COUNTRY ANALYSIS

Abstract

This study applies Long Short-Term Memory (LSTM) neural networks to forecast asset prices within green and non-green cryptocurrency portfolios across ten geographically diverse markets: the United States, the United Kingdom, Canada, Germany, France, Italy, Japan, India, Russia, Brazil, South Africa, and China. The LSTM architecture is selected for its demonstrated capacity to capture long-range temporal dependencies within financial time series data, overcoming the vanishing gradient limitations of conventional recurrent networks. Employing daily closing price data for the period ending June 2024, the research constructs and evaluates portfolios that blend cryptocurrencies with equities, commodities, and sector-specific instruments within each national context. Results indicate that LSTM models achieve moderate accuracy for stable conventional equities and select commodities, while systematically overestimating or underestimating highly volatile cryptocurrency assets. Notable cross-regional heterogeneity is observed, with BRICS economies displaying wider prediction divergences compared to G7 counterparts.