PaperlayerView exact revision on arXiv

Source-linked AI summary

A Gated Recurrent Unit Approach to Bitcoin Price Prediction

Aniruddha Dutta, Saket Kumar, Meheli Basu

arXiv:1912.11166v1q-fin.PRq-fin.MF

TL;DR

Bitcoin-price studies have often relied on limited feature sets, while sequence models with robust feature engineering have remained uncommon. This study engineers exogenous and endogenous features and compares prediction models, finding that GRU performed better than LSTM and recurrent dropout improved GRU performance.

  • Problem

    Prior Bitcoin-price studies often relied on pre-decided models with limited features, while GRU models had received little attention in cryptocurrency price prediction.

    Source: §1 · p. 2, §6 · p. 11

  • Method

    The study engineers exogenous and endogenous Bitcoin-price features and compares time-series and machine-learning models, including a proposed GRU architecture.

    Source: §1 · p. 2, §3 · p. 4, §3.1 · p. 5

  • Results

    LSTM and GRU outperformed conventional time-series approaches such as SARIMA, while GRU performed better than LSTM and recurrent dropout improved GRU performance.

    Source: §6 · p. 12

  • Takeaways & Limitations

    The authors conclude that trained GRU models may benefit investment strategies through financial gains and portfolio management.

    Source: §6 · p. 12

    Preprocessing assumes a midnight Bitcoin end-of-day price and carries Friday stock, bond, and commodity prices through weekends, ignoring after-market trading.

    Source: §3.1 · p. 5

Abstract

from arXiv · show

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. Introduction

The paper proposes a GRU-based approach to Bitcoin price prediction using exogenous and endogenous variables, comparing it with other models and testing trading strategies. It addresses a forecasting challenge shaped by high volatility and gaps in prior cryptocurrency machine-learning research.

1. Introduction

  • High volatility makes cryptocurrency forecasting distinct from traditional financial-market forecasting, while traditional time-series methods are described as ineffective amid volatility and absent seasonality.The paper frames accurate price forecasting as a challenge for investors and government agencies.Source: §1 · pp. 1–2
  • Prior Bitcoin studies have used random forests, Bayesian neural networks, RNNs, and LSTMs; reported comparisons found RNN/LSTM and LSTM outperforming MLP and ARIMA, respectively (Madan 2015; Jang 2017; McNally 2018; Guo 2018; Karakoyun 2018).The introduction identifies a dearth of machine-learning applications in cryptocurrency price prediction and presents nonlinear modeling as relevant to volatile crypto prices.Source: §1 · p. 2
  • The study combines exogenous and endogenous variables, proposes a GRU architecture for Bitcoin price prediction, and tests two simple trading strategies.The authors say some variables have not been investigated in previous Bitcoin price-prediction studies, and that the GRU has not previously been investigated for cryptocurrency price prediction.Source: §1 · p. 2
  • The authors compare the proposed GRU with traditional machine-learning, neural-network, and time-series models across different lookback periods to assess robustness.They state that GRUs train faster than traditional RNNs or LSTMs and that their architecture can learn Bitcoin price fluctuations more efficiently than traditional LSTM.Source: §1 · p. 2

2. Methodology

The methodology compares NN, LSTM, and GRU models for Bitcoin price prediction against SARIMA, and uses the best-performing model for portfolio strategy execution. It explains how LSTM and GRU gates manage sequence information and address vanishing gradients.

  • Traditional RNNs can lose long-range sequence information because of short-term memory and vanishing gradients (Wang, 2015; Sepp Hochreiter, 1998, Razvan Pascano 2013).Source: §2 · p. 2
  • LSTMs are designed to learn long-term dependencies, using forget, input, and output gates to regulate information flow through the cell and hidden state (Sepp Hochreiter, 1997).Source: §2 · p. 3
  • GRUs combine LSTM’s forget and input gates into an update gate, merge cell and hidden states, and use update and reset gates to control what passes forward or is forgotten.The update gate addresses vanishing gradients; the reset gate determines how much past information to forget, while the update gate balances current content with past information.Source: §2 · pp. 3–4
  • Bitcoin-price experiments compare optimized, test-set NN, LSTM, and GRU models with SARIMA, then use the best-performing model for portfolio strategy execution.Source: §2 · p. 4
  • GRUs use fewer tensor operations and train faster than LSTMs (Junyoung Chung, 2014).Source: §2 · p. 4

3. Data collection and Feature Engineering

The study assembled endogenous Bitcoin measures and exogenous economic indicators for price prediction, then standardized and split the data and selected features using correlation analysis.

3. Data collection and Feature Engineering

  • Predictors combined Bitcoin market and technical measures with economic and financial indicators, with MACD engineered to explore moving-average signals.Endogenous features included price, lag returns, volatility, miner revenue, transaction volume and fees, hash rate, money supply, block size and Metcalfe-UTXO; exogenous features included Treasury yields, gold, VIX and S&P dollar returns.Source: §3 · pp. 4–5

3.1. Data pre-processing

  • Data were normalized by demeaning each series and dividing by its standard deviation, then split chronologically into training, validation and test periods.End-of-day Bitcoin price was set at midnight; stock, bond and commodity prices were assumed to remain at Friday’s level on weekends. The periods were Jan 1, 2010–June 30, 2018; July 1–Dec 31, 2018; and Jan 1–June 30, 2019, respectively. Lookbacks of 15, 30, 45 and 60 days predicted the next day’s price.Source: §3.1 · pp. 5–6

3.2. Feature Selection

  • Spearman correlation analysis excluded features with Bitcoin-price correlation coefficients of 0.8 or greater, leaving 15 features.The five dropped features were Bitcoin Miner Revenue, Metcalf-UTXO, Interest rates, Block size and US Bond yields 2-years and 10-years difference.Source: §3.2 · pp. 6–7, Figure 5 · p. 7

4. Model Implementation and Results

The study finds that a GRU with recurrent dropout is the best-performing model for Bitcoin price prediction, with a 30-day lookback producing the best RMSE results.

  • The study compared SARIMA and neural models, including LSTM and GRU variants, and preferred RMSE over MAE because it weights large errors more heavily.The candidate neural models included dropout and recurrent-dropout variants; the study also used training and validation losses to assess bias and variance.Source: §4 · pp. 7–8
  • Models were optimized with hyperparameter tuning; TanH was selected by test RMSE, Adam was used, and changing the learning rate yielded little improvement.RNN lookback periods were tested at 15, 30, 45, and 60 days; neural-network and RNN batch sizes were 125 and 100, respectively.Source: §4 · pp. 8–9
  • The GRU with recurrent dropout achieved RMSE 0.014 on training data and 0.017 on test data, and was the best-performing model overall.The study reports that recurrent dropout reduced test RMSE and gave this model the smallest train–test RMSE difference.Source: §4 · p. 9, Table 1 · p. 9
  • A 30-day lookback period produced the best RMSE results among the tested 15-, 30-, 45-, and 60-day periods.The 30-day period was initially selected to align with a standard monthly trading calendar.Source: §4 · pp. 8–9, Table 2 · p. 10
  • With recurrent dropout, the one-layer GRU’s training–validation loss gap narrowed, but it widened after adding a second GRU layer.Figure 6 compares training and validation losses for the simple NN, LSTM, and GRU variants.Source: Figure 6 · p. 8, §4 · p. 9
  • The GRU with dropout and recurrent dropout predicted prices above the original early in 2019, then learned the upward trend during the June–July price rise.The passage describes this comparison against original test-data prices.Source: §4 · p. 9, Figure 7 · p. 10

5. Portfolio Strategy

The section evaluates two signal-driven Bitcoin trading strategies using the one-layer GRU with recurrent dropout. On test data, buy-sell handled the June 2019 price fall better than long-short, whose daily settlement fees dampened growth and increased volatility.

  • The test-set strategy uses the one-layer GRU with recurrent dropout: long-short trades daily signals (0.8% transaction fees, initial value 1), whereas buy-sell stays invested until a sell signal.Source: §5 · p. 10
  • Figure 8 shows the buy-sell strategy avoiding the June 2019 fall by moving to cash, while long-short growth remained slow; daily settlement fees reduced its growth and increased volatility.Source: §5 · p. 11
  • Most Bitcoin exchanges did not yet allow short selling (Primavera De Filippi, 2014); the paper suggests long-short may be more useful across cryptocurrencies with simultaneous long and short positions.Source: §5 · p. 11

6. Conclusions

The study combines finance-linked feature engineering with recurrent models for Bitcoin price prediction, highlighting GRU with recurrent dropout. It also reports potential benefits from applying trained GRU models to investment strategies.

6. Conclusions

  • The study uses a holistic, finance-linked predictor set to learn future Bitcoin price trends.Features include US Treasury 2-year and 10-year yields, Google Trends, VIX, MACD, and lagged prices.Source: §6 · p. 11
  • LSTM and GRU models outperform conventional time-series approaches such as SARIMA for Bitcoin price prediction.The conclusion notes that market data is limited, allowing neural networks to learn nonlinear patterns, though they require accurate training and hyperparameter tuning.Source: §6 · p. 12
  • The authors note that deep-model training and tuning can be computationally extensive, while a larger dataset could improve the study.They consider computational complexity less concerning for stock and cryptocurrency prediction because market data is limited.Source: §6 · p. 12
  • GRU performs better than LSTM in the study, and recurrent dropout improves GRU performance.The authors attribute GRU’s performance to its simultaneous forgetting and updating, and call for further study of recurrent dropout.Source: §6 · p. 12
  • The authors report that two investment strategies implemented with the trained GRU can benefit investment practice through financial gains and portfolio management.They say proper backtesting could help manage portfolio risk and reduce financial losses.Source: §6 · p. 12

Appendix A · Definition of variables and data source

Appendix A defines Bitcoin-network, market, and macroeconomic variables used in the study, including measures of network activity, market expectations, and external indicators.

Definition of variables and data source

Related papers

Loading 1912.11166v1…