| Résumé : |
More and more corporations as well as individual and institutional investors starting to invest in Bitcoin (and other cryptocurrencies), resulting in price increases and more legitimacy for this new asset class. Investment decisions always beg the question about future price development and the knowledge about it supports decisions regarding portfolio optimization, risk evaluation and trading. Since financial time series are usually highly nonlinear and noisy, it calls for alternatives to classical time series analy- sis. The emergence of advanced machine learning systems attracted the attention of professionals and academics, utilizing them successfully to model and predict the prices of stocks and cryptocurrencies. Most successful are deep learning (DL) artificial neural networks (ANNs) with long-short term memory (LSTM). Therefore, with this dissertation, LSTM ANNs were exploited to predict Bitcoin’s daily price, using a dataset from 01/01/2015 to 31/05/2021. To give the reader a basis and better understanding firstly the backgrounds of cryptocurrencies and ANNs were introduced. Afterwards, eight different training algorithms including, SGD, RMSprop, Adam, Adadelta, Adagrad, Adamax, Nadam, and FTRL, using five different epoch sizes including, 25, 50, 100, 250 and 500 were used to build up LSTM ANNs. By evaluating the forecasting performance using the root mean square error (RMSE), it was confirmed that the choice of the training algorithm and the epoch size significantly influences the forecasting performance of Bitcoin’s daily price. Thereby, the FTRL was identified as completely unsuitable, and the Adamax using 25 epochs as the optimal algorithm, even in times of high volatility. To the best of the author’s knowledge, this was the first time that this was investigated on an empirical basis. |
Ajouter le résultat dans votre panier Faire une suggestion Affiner la rechercheCorrelation between nft & crypto assets And different asset classes & impact on Portfolio diversification / 2022
L'art du trading / EYROLLES (2015)
