يونس محمودي

Forecasting financial markets has perpetually remained a formidable challenge in econometrics. With the advent of cryptocurrencies, the highly volatile an‎d noisy nature of these assets has exposed the limitations of conventional forecasting approaches. The primary objective of this research is to propose an accurate an‎d robust forecasting framework based on the "accuracy versus computational cost" trade-off, utilizing the Long Short-Term Memory (LSTM) family an‎d their systematic optimization. In this thesis, an integrated 5 year hourly frequency dataset comprising seven leading crypto assets was extracted. To prevent data leakage an‎d simulate real world trading conditions, a rigorous 1 step ahead rolling forecast eva‎luation was conducted across 11 diverse architectures spanning statistical methods, tree based algorithms, recurrent networks, convolutional models, transformers such as TFT an‎d N-BEATS, an‎d foundation language models such as Chronos. The empirical findings indicate that in noisy financial environments, linear statistical models such as ARIMA perform exceptionally well, serving as an empirical validation of the ran‎dom walk hypothesis. Conversely, tree based models fail to extrapolate during historical breakouts. Although massive transformer architectures achieved the highest precision, recording an R-squared value of 0.9990 for Bitcoin, their computational overhead an‎d inference latency are disproportionately high. Meanwhile, by calibrating the hyperparameter space via Grid Search, the optimized Bidirectional LSTM secured an exceptionally close performance, recording a Mean Absolute Percentage Error of 0.56 percent an‎d an R-squared value of 0.9976, at only a fraction of the training time an‎d computational cost. This achievement demonstrates that recurrent networks, owing to their Markovian inductive bias, establish the Pareto optimal frontier between prediction accuracy an‎d inference speed, remaining the most pragmatic engineering choice for developing high frequency algorithmic trading systems.

ارزهاي ديجيتال , يادگيري عميق , شبكه‌هاي بازگشتي , ترنسفورمرها , جستجوي شبكه‌اي , موازنه هزينه-دقت , ال‌اس‌تي‌ام

Cryptocurrencies , Deep Learning , Recurrent Neural Network , Transformers , Grid Search , Accuracy-Efficiency Trade-off , LSTM

Younes Mahmoudi

Mohammadhadi Alaeiyan