Evaluating the Performance of a Hybrid Model in Improving Key Risk Management Metrics in Financial Markets
TL;DR
The hybrid model significantly outperformed all benchmark models, achieving the lowest mean absolute error, root mean squared error, and mean absolute percentage error, as well as the highest directional accuracy, F1 score, and area under the curve.
Abstract
This study aimed to design and evaluate a hybrid financial–behavioral model integrating technical market indicators, volatility features, and social-media sentiment to improve key risk-management metrics in Bitcoin and Ethereum markets. This quantitative, ex post facto, developmental–applied study used hourly and daily financial data for Bitcoin and Ethereum from January 2019 to December 2025, together with English-language posts from Twitter/X containing relevant cryptocurrency keywords and hashtags. Financial variables included prices, trading volume, returns, volatility measures, and more than 50 technical indicators. Textual data were processed using natural language processing procedures, and sentiment scores were extracted through a finance-specific BERT model. The proposed FinBERT–LSTM model combined financial, technical, and behavioral features. Its performance was compared with naïve persistence, ARIMA, GARCH, random forest, financial LSTM, and sentiment-only models using chronological train–validation–test partitions and walk-forward validation. Predictive accuracy, directional classification, value-at-risk calibration, maximum drawdown, expected shortfall, and risk-adjusted performance were evaluated. The hybrid model significantly outperformed all benchmark models, achieving the lowest mean absolute error, root mean squared error, and mean absolute percentage error, as well as the highest directional accuracy, F1 score, and area under the curve. Diebold–Mariano tests confirmed significantly lower forecasting errors than the financial LSTM and sentiment-only models (p < 0.001). The hybrid strategy also produced lower annualized volatility, downside deviation, maximum drawdown, value at risk, and expected shortfall, while yielding higher Sharpe, Sortino, and Calmar ratios. Kupiec and Christoffersen tests indicated adequate value-at-risk coverage and independence at the 95% and 99% confidence levels. Ablation analyses further showed that removing sentiment, technical, engagement, or volatility features significantly weakened predictive and risk-management performance. Integrating technical, temporal, and behavioral information within a hybrid deep-learning architecture improves both forecasting accuracy and the management of downside and tail risk in cryptocurrency markets.