Speeches di Matteo Testi

Volatility: your Enemy or your Friend? A Deep Learning Solution

Cryptocurrency time series are usually known to be very complex, non-stationary and very noisy. Furthermore, in recent years, more and more Deep Learning models have been employed to “Sentiment Analysis” (SA) thanks to their automatic high-dimensional feature extraction capability even from non-structured symbolic data such as generic texts (i.e. tweet). These breakthroughs are also due to the latest pre-processing discovers like Word Embeddings (Mikolov, 2013). In this work we propose a novel stack of cutting-edge Deep Learning models aimed to generate outperforming prediction accuracy and profitability in cryptocurrency forecasting. Namely, we leverage a new Deep Learning time-series model called WSAEs-LSTM by efficiently combining Wavelet-Trasforms (WT) for data denoising; Stacked Autoencoders (SAE) for high-level feature extraction and LSTM (Long Short Term Memory) for modelling cryptocurrency prices historical data. On the other side, a Recurrent Convolutional Neural Network model (RCNN) achieving the “state-of-art” accuracy in sentiment classification, has been implemented (Lai, et al., 2015). Our RCNN implementation achieves 85% of accuracy on this training set. Using RCNN we produce a sentiment time-series for each textual sequence which is eventually inserted as input to WSAEs-LSTM. The high accuracy of RCNNs and WSAEs-LSTM enables us to reach supreme levels of crypto-coin predictions by measuring RMSE (and MAPE) between real values of “close” and “volume” attribute and the predicted ones, as far as the -test set is exclusively concerned. The historical data entered into WSAEs-LSTM (other than the Sentiment time-series) are “high”, “low”, “open”, “close”, “volume”, “weighted average” attributes of 63 different cryptocurrencies.

Topics

Industry 4.0, Fintech/blockchain, Social media analytics, Predictive analysis, Algorithms

All Speakers