Source-linked AI summary
A Gated Recurrent Unit Approach to Bitcoin Price Prediction
Aniruddha Dutta, Saket Kumar, Meheli Basu
TL;DR
Bitcoin’s high volatility and limited prior use of sequence models with robust feature engineering motivate a daily price-prediction framework. The paper combines exogenous and endogenous factors, compares neural and time-series models using RMSE, and evaluates trading strategies. GRU with recurrent dropout performs best in the reported analysis, while the authors conclude that such models can support financial gains and portfolio management when properly implemented.
Problem
Bitcoin’s high volatility and limited prior use of sequence models with robust feature engineering leave a need for accurate price forecasting.
Method
The study engineers exogenous and endogenous features, compares NN, LSTM, GRU, and SARIMA models using RMSE, and evaluates two Bitcoin trading strategies.
Results
GRU with recurrent dropout performs better than the LSTM model in the authors’ analysis and is selected as the best model for trading-strategy evaluation.
Takeaways & Limitations
The authors report that properly implemented machine-learning models can benefit financial gains and cryptocurrency portfolio management.
Takeaways & Limitations
The study assumes weekend stock, bond, and commodity prices remain at Friday’s values while cryptocurrency trades continuously.
Abstract
from arXiv · showhide
In today's era of big data, deep learning and artificial intelligence have formed the backbone for cryptocurrency portfolio optimization. Researchers have investigated various state of the art machine learning models to predict Bitcoin price and volatility. Machine learning models like recurrent neural network (RNN) and long short-term memory (LSTM) have been shown to perform better than traditional time series models in cryptocurrency price prediction. However, very few studies have applied sequence models with robust feature engineering to predict future pricing. in this study, we investigate a framework with a set of advanced machine learning methods with a fixed set of exogenous and endogenous factors to predict daily Bitcoin prices. We study and compare different approaches using the root mean squared error (RMSE). Experimental results show that gated recurring unit (GRU) model with recurrent dropout performs better better than popular existing models. We also show that simple trading strategies, when implemented with our proposed GRU model and with proper learning, can lead to financial gain.
1 Haas School of Business, University of California Berkeley, CA 94720, USA
The passage lists keywords covering Bitcoin, artificial intelligence, neural networks, time-series analysis, predictive modeling, trading strategies, and risk management.
- The paper concerns Bitcoin and cryptocurrency applications of artificial intelligence and deep learning.
- Its methods and topics include neural networks, predictive models, and time series analysis.
- The application areas include trading strategy and risk management.
1. Introduction
Bitcoin’s high volatility creates a need for accurate price forecasting, while prior work has relied heavily on traditional models or limited machine-learning applications. The paper addresses this gap by combining exogenous and endogenous variables, comparing model classes, and testing trading strategies.
- Bitcoin’s high volatility and fluctuating price trends create a need for standardized and accurate cryptocurrency forecasting.
- Traditional time-series methods such as AR, MA, SES, and ARIMA are described as ineffective for highly volatile cryptocurrency markets lacking seasonality.
- Prior Bitcoin prediction studies include random forests, Bayesian neural networks, and neural networks, but machine-learning applications remain limited.
- The paper combines exogenous and endogenous variables and compares RNN models with traditional machine-learning and time-series models.
- The proposed GRU architecture is evaluated across different lookback periods and used to test two simple algorithmic-trading strategies.
2. Methodology
The methodology reviews sequence models and develops a GRU-based forecasting approach alongside neural-network and SARIMA baselines. GRU gates regulate past and current information, with recurrent dropout considered during model evaluation.
- RNNs process sequences with internal memory but can fail to learn long-term dependencies because of vanishing gradients.
- LSTMs use input, cell, forget, and output gates to regulate information flow and learn long-term dependencies.
- GRUs use update and reset gates to control how much current and past information enters the final memory state.
- GRUs require fewer tensor operations and train faster than LSTMs, motivating their inclusion in Bitcoin price prediction.
- The evaluated models include a simple neural network, LSTM, GRU, and SARIMA, with optimized hyperparameters and test-set evaluation.
3. Data collection and Feature Engineering
The study engineers Bitcoin-price features from endogenous cryptocurrency data and exogenous economic indicators, then normalizes and splits the data for forecasting. Correlation analysis is used to reduce the feature set before modeling.
- The feature set combines Bitcoin price, lagged returns, volatility, mining, transaction, network, monetary, and broader financial indicators.
- The data are normalized after merging sources, with observations split into training, validation, and test periods from 2010 through June 2019.
- Lookback periods of 15, 30, 45, and 60 days are used to predict the following one-day Bitcoin price.
- Endogenous features are more correlated with Bitcoin prices than exogenous features, while Google trends, interest rates, and Ripple price are prominent exogenous correlates.
- Correlation analysis selects 15 features after removing Bitcoin Miner Revenue, Metcalf-UTXO, interest rates, block size, and specified US bond-yield differences.
4. Model Implementation and Results
The study compares traditional time-series and neural-network approaches for Bitcoin forecasting, including dropout variants of LSTM and GRU models. The GRU with recurrent dropout achieves the reported strongest RMSE performance among the evaluated models.
- Modeling approach: SARIMA, a simple neural network, LSTM, GRU, and GRU dropout variants are trained and evaluated for Bitcoin price forecasting.The study uses RMSE for model-error evaluation and considers SARIMA because the Bitcoin price series appears non-stationary with seasonality and trend.
- Regularization: Dropout and recurrent dropout are introduced in LSTM and GRU layers to reduce complex co-adaptations and overfitting.The recurrent dropout rate is 0.1 for both LSTM and GRU architectures; the two-hidden-layer GRU additionally uses dropout of 0.1.
- Architecture: The final GRU architecture uses two GRU layers with 50 and 10 nodes, plus dropout and recurrent dropout rates of 0.1.The simpler GRU and recurrent-dropout GRU use one 50-node GRU layer and one one-node dense layer.
- Results: 0.014 training RMSE and 0.017 test RMSE are reported for the GRU model with recurrent dropout.These values are reported for the model’s train and test data, respectively.
- Results: The training-validation loss difference decreases for the one-layer GRU with dropout and recurrent dropout but increases after adding another GRU layer.Figure 6 compares training and validation loss across the neural-network architectures.
5. Portfolio Strategy
The paper evaluates the selected one-layer GRU with recurrent dropout through long-short and buy-sell Bitcoin trading strategies. The results show that transaction costs and short-selling exposure materially affect portfolio behavior.
- Strategy design: Two Bitcoin trading strategies are implemented using signals from the GRU one-layer model with recurrent dropout.The long-short strategy buys or short-sells based on predicted daily price direction, while the second strategy uses buy and sell signals.
- Long-short strategy: The long-short portfolio incurs transaction fees from daily settlement, reducing growth and increasing portfolio volatility.The strategy assumes transaction fees of 0.8% of the invested or sold amount.
- Observed test-period behavior: During Bitcoin’s early-June 2019 decline, the buy-sell strategy correctly predicts the fall and preserves cash before reinvesting as prices rise.The long-short strategy increases portfolio value more slowly during the same period because it combines short and long positions.
6. Conclusions
The study concludes that recurrent neural networks, particularly GRU models, outperform conventional SARIMA for Bitcoin price prediction. It also reports that recurrent dropout improves GRU performance and that model-based strategies can support portfolio management, while larger datasets could improve the study.
- Conclusions: The study uses an extensive feature set with financial linkages to investigate Bitcoin price prediction and its potential as an alternative investment.The features are selected to capture possible diversification benefits and links with broader financial conditions.
- Conclusions: LSTM and GRU models outperform conventional time-series approaches such as SARIMA for Bitcoin price prediction.The authors attribute the result in part to the neural networks’ ability to learn non-linear patterns from limited data.
- Conclusions: GRU performs better than LSTM in this analysis, and adding recurrent dropout improves the GRU architecture’s performance.The authors state that further studies are needed to explore dropout behavior in GRU architectures.
- Implications: Machine-learning trading strategies can benefit investment and portfolio management when implemented with full understanding and proper backtesting.The authors state that proper backtesting can contribute to managing portfolio risk and reducing financial losses.
- Scope and future work: A bigger dataset is identified as a potential source of significant improvement over the current study.The conclusion also notes that deep models require accurate training and hyperparameter tuning, which can be computationally extensive for large datasets.
Definition of variables and data source
The study defines endogenous Bitcoin-network variables and exogenous financial and market indicators used as predictive features, drawing data from multiple sources.
- Hash rate measures the number of block solutions computed per second by all miners.
- Money supply represents the amount of Bitcoin Core (BTC) in circulation.
- Metcalfe-UTXO applies Metcalfe's Law, linking network value to the square of the number of participants.
- Block size measures transaction data organized into cryptographically linked blocks, which increase as network use grows.
- Exogenous indicators include Google Trends, VIX volatility, gold prices, the U.S. Dollar Index, bond yields, MACD, Ripple prices, stock-market returns, and interest rates.VIX represents expected 30-day forward-looking volatility, while the U.S. Dollar Index measures the dollar against a basket of major trading-partner currencies.