Cryptocurrencies have experienced significant growth in recent years, motivating the need for robust models to forecast price movements. This study employs a Long Short-Term Memory (LSTM) network to predict the directional movement of the daily closing price of the Bitcoin/USD currency pair. An experimental evaluation is carried out to assess both the predictive accuracy of the proposed model and the financial outcomes—profit or loss—of trades executed based on its forecasts. Additionally, we analyze how the input-window dimensionality and the Largest Lyapunov Exponent (LLE) affect forecasting performance. The research also explores the potential risks and actual threats that traders may face when relying on the model's predictions for trading decisions. Bayesian optimization is used to fine-tune the LSTM parameters, which leads to a notable improvement in model accuracy. Furthermore, analysis of metrics such as drawdown and reward-to-risk ratios from model-driven trades reveals that the model supports confident trading, thanks to the limited exposure to both realized and anticipated risks.

This study (i) incorporates a chaos-based feature, the Largest Lyapunov Exponent (LLE), into an LSTM framework for next-day directional forecasting of BTC/USD, (ii) applies Bayesian optimization to tune LSTM hyperparameters, (iii) evaluates practical tradability via profitability and risk measures (cumulative return, drawdown, reward-to-risk), and (iv) investigates the impact of input-window dimensionality on predictive and trading performance.