Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: M.Sc. Thesis
- Language: English
- Document No: 55814 (58)
- University: Sharif University of Technology, International Campus, Kish Island
- Department: Science and Engineering
- Advisor(s): Khayyat, Amir Ali Akbar
- Abstract:
- Cryptocurrencies are not only regarded as a trustworthy method of financial transaction validated by a decentralized cryptographic system as opposed to a centralized authority, but also as one of the most popular and lucrative forms of trade and investing. Predicting the price of a cryptocurrency is a challenging topic in time-series research. Its intricacy is due to the volatility and large swings of cryptocurrencies' price. The emergence of brand-new cryptocurrencies, which might present a profitable trading opportunity but lack sufficient historical data for technical analysis, prompted us to develop a trading strategy that could be applied universally. The forecast of the next timestep's price, while essential for a trading system, is insufficient. On top of that, we require a decision-making core that should generate buy and sell signals while being fed with predictions that forecast one or more steps ahead of time. We have proposed a novel approach for price prediction using technical analysis: a describing function that would map the price trend to the time series of technical rule signals through utilizing technical indicators and strategies and inputs them to four deep learning methods: simple LSTM, encoder-decoder LSTM with single attention mechanism, encoder-decoder LSTM with dual attention mechanism, and our proposed deep learning architecture, which is an encoder-decoder convolutional transformer to predict the percentage of price changes. In doing so, not only could we mitigate the information leakage issue but also establish the groundwork for generalizable trading systems. Furthermore, we have proposed a novel decision-making system in which we have employed an autoencoder to classify sell and buy positions while being provided with the aforementioned technical descriptor signals as its inputs. Having tested on bitcoin, the auto-encoder decision maker managed to generate buy and sell orders, resulting in the identification of two-thirds of all possible profits from price fluctuation over the unseen testing data and 56 percent of all possible profits even when it is trained on an asset other than the one tested, highlighting our foundational fact that the price trends of different assets share similar technical patterns. Incorporating the autoencoder core with deep learning price predictor networks would form a robust hybrid trading system. Regarding the deep learning time series predictors, when it comes to feeding our autoencoder decision maker with predictions made one step ahead of time, deep learning systems equipped with multi-attention mechanisms are more likely to meet our highest performance expectations. This is because the predictor system must identify the most effective signals while selecting the most relevant time periods. Moreover, the results of testing our hybrid trading system on bitcoin indicate that the autoencoder that has learned the technical pattern would significantly filter out the false signals generated by the deep learning predictor models if used as a standalone trading system that obeys its own predictions for price direction as the trading signals and significantly increases profits. Our proposed encoder-decoder convolutional transformer outperforms the other tested models both as a standalone trading system and as a hybrid one integrated with our proposed autoencoder trading decision maker system, which is capable of capturing approximately 20% and 40% of all possible profit as a standalone and hybrid model, respectively
- Keywords:
- Artificial Intelligence ; Machine Learning ; Time Series ; Attention Mechanism ; Cryptocurrency ; Long Short Term Memory (LSTM) ; Bitcoin ; Technical Analysis
Digital Object List-
محتواي کتاب - view
Bookmark