PaperlayerView exact revision on arXiv

Source-linked AI summary

Predicting and Forecasting the Price of Constituents and Index of Cryptocurrency Using Machine Learning

Reaz Chowdhury, M. Arifur Rahman, M. Sohel Rahman, M. R. C. Mahdy

arXiv:1905.08444v1q-fin.ST

TL;DR

Cryptocurrency's high volatility makes price prediction challenging, and the literature on price fluctuations is described as limited. This paper applies four machine-learning models to close prices for CCI30 and nine cryptocurrencies; the authors report good overall prediction performance and competitive comparisons with prior work, while K-NN forecasting was less effective.

  • Problem

    Cryptocurrency's extreme volatility complicates price-fluctuation prediction, and the authors describe relevant research as limited.

    Source: §1 · p. 3, §2 · p. 4

  • Method

    The paper applies gradient boosted trees, a neural network, ensemble learning, and K-NN to predict or forecast close prices for CCI30 and nine cryptocurrencies.

    Source: §1 · p. 4, §3.3.1 · p. 6, §3.3.2 · p. 7

  • Results

    The authors report good overall close-price prediction and competitive performance against prior work, but say K-NN forecasting was not effective.

    Source: §4 · p. 10

  • Takeaways & Limitations

    The models are presented as useful for understanding cryptocurrency trends and patterns in the tested close-price setting.

    Source: §4 · p. 10

    K-NN forecasting was less effective, which the authors attribute to noisy random features and extreme volatility.

    Source: §4 · p. 10

Abstract

from arXiv · show

At present, cryptocurrencies have become a global phenomenon in financial sectors as it is one of the most traded financial instruments worldwide. Cryptocurrency is not only one of the most complicated and abstruse fields among financial instruments, but it is also deemed as a perplexing problem in finance due to its high volatility. This paper makes an attempt to apply machine learning techniques on the index and constituents of cryptocurrency with a goal to predict and forecast prices thereof. In particular, the purpose of this paper is to predict and forecast the close (closing) price of the cryptocurrency index 30 and nine constituents of cryptocurrencies using machine learning algorithms and models so that, it becomes easier for people to trade these currencies. We have used several machine learning techniques and algorithms and compared the models with each other to get the best output. We believe that our work will help reduce the challenges and difficulties faced by people, who invest in cryptocurrencies. Moreover, the obtained results can play a major role in cryptocurrency portfolio management and in observing the fluctuations in the prices of constituents of cryptocurrency market. We have also compared our approach with similar state of the art works from the literature, where machine learning approaches are considered for predicting and forecasting the prices of these currencies. In the sequel, we have found that our best approach presents better and competitive results than the best works from the literature thereby advancing the state of the art. Using such prediction and forecasting methods, people can easily understand the trend and it would be even easier for them to trade in a difficult and challenging financial instrument like cryptocurrency.

1. Introduction

Cryptocurrency volatility makes price and risk assessment difficult, despite growing investment and prior modeling efforts. This paper compares machine-learning models for forecasting close prices of CCI30 and nine cryptocurrencies.

1. Introduction

  • High volatility and sharp price swings make cryptocurrency risk assessment difficult and can lead to rapid devaluation.The paper also describes these currencies as decentralized and unregulated.Source: §1 · p. 3
  • Bitcoin's blockchain is described as a public, peer-to-peer ledger for recording cryptocurrency transactions.The paper notes its role in addressing double spending and reducing centralized control over value transfers.Source: §1 · p. 3
  • Earlier cryptocurrency research used methods including machine learning, data mining, text mining, agent-based markets, genetic programming, and correlation analysis.Prior work also examined machine learning for abnormal-profit strategies and measuring illicit activity [12].Source: §1 · p. 3
  • The study applies ensemble learning, gradient boosted trees, neural networks, and K-NN to historical close prices for CCI30 and nine cryptocurrencies.The data were collected from CoinMarketCap, and the models were developed using RapidMiner.Source: §1 · p. 3
  • The authors present the models as tools for observing cryptocurrency behavior and assessing which algorithms may better predict or forecast close prices.Source: §1 · p. 4

2. Related Works

The reviewed literature spans statistical, machine-learning, deep-learning, and transaction- or community-based approaches to cryptocurrency prediction. The authors characterize the literature as limited given the market's high volatility.

2. Related Works

  • The authors describe cryptocurrency price-prediction research as limited, given the asset's high volatility and price fluctuations.Source: §2 · p. 4
  • LightGBM was reported to outperform SVM and random forest for cryptocurrency price-trend forecasting, with accuracy varying across training-set categories and forecast horizons.Reported accuracies include 0.905 for two weeks in one category and 0.952 for two weeks in the other.Source: §2 · p. 4
  • For Bitcoin, Digital Cash, and Ripple, DLNN and GRNN produced markedly different RMSE values, with DLNN lower than GRNN for all three currencies.The reported DLNN and GRNN RMSE pairs are 2.75 × 10^3 and 8.80 × 10^3 for Bitcoin, 19.2923 and 50.2418 for Digital Cash, and 0.0499 and 0.3115 for Ripple.Source: §2 · p. 4
  • A stacking study combined classifiers to predict Bitcoin price direction across April–May and June–July 2018 periods.Its reported accuracies vary by model and period; the stacked generalization values were 0.52 and 0.54, respectively.Source: §2 · pp. 4–5
  • RNN and LSTM outperformed ARIMA in one Bitcoin price-direction study, while another study reported ARIMA MAPE of 5.36% across September and October 2017.In the first study, LSTM and RNN accuracy/RMSE were 52.78%/6.87% and 50.25%/5.45%, versus ARIMA's 50.05%/53.74%.Source: §2 · p. 5
  • Other studies used seq2seq and ARIMA models, online-community comments, and Bitcoin transaction graphs to predict prices or fluctuations.The seq2seq and ARIMA predictions were described as substantially deviating from Bitcoin's true values over 40 days.Source: §2 · pp. 4–5

3. Materials and Methods

The study uses daily CoinMarketCap data to predict January 2019 close prices for CCI30 and its constituents. It evaluates models with regression and forecasting measures.

3.1 Dataset

  • Seven-day-week daily data from CoinMarketCap were split into training and testing sets to predict January 2019 close prices for the cryptocurrencies and CCI30.K-NN uses training data only, because the authors use it to forecast January values.Source: §3.1 · p. 5

3.2 Performance Metrics

  • The evaluation measures include RMSE, prediction-trend accuracy, absolute and relative error, squared error, correlation, and squared correlation.Source: §3.2 · pp. 5–6
  • Prediction-trend accuracy measures how often a regression prediction correctly identifies the trend; relative error averages absolute deviation divided by the actual value.RMSE measures differences between predicted and actual values; correlation and squared correlation compare the label and prediction attributes.Source: §3.2 · pp. 5–6

3.3 Methodology

The methodology compares four machine-learning models for cryptocurrency close-price prediction and forecasting. Their implementations use historical data and distinct training, windowing, and voting configurations.

3.3 Methodology

  • The authors use RapidMiner to train models on datasets and visualize or compare their outputs, motivated by cryptocurrency price volatility.They describe machine learning as relevant to finance applications including portfolio management and price-fluctuation prediction.Source: §3.3 · p. 6

3.3.1 Predictive Model: Gradient Boosted Trees

  • Gradient boosted trees combine regression or classification trees through forward learning; the January 2019 model uses 500 trees, with other parameters left at defaults.The authors describe the method as scalable and rule-based, and note that it does not require data normalization.Source: §3.3.1 · p. 6

3.3.2 Predictive Model: Neural Net

  • The neural-net model uses windowed historical inputs to predict future close prices, with a four-example training window and 500 training cycles.The window size and step size are both 1; learning rate is 0.03 and momentum is 0.9.Source: §3.3.2 · p. 7

3.3.3 Predictive Model: Ensemble Learning Method

  • The ensemble splits data 60/40 and averages predictions from gradient boosted trees, a neural net, and linear regression within a voting operator.The neural net uses 500 cycles, learning rate 0.3, and momentum 0.2; the gradient boosted trees model uses 500 trees.Source: §3.3.3 · pp. 7–8

3.3.4 Predictive Model: K-NN

  • K-NN compares an unknown example with its nearest training examples; for forecasting, the model uses only Date and Close Price attributes.The authors forecast January 2019 and assess residuals, which they describe as ideally uncorrelated and with a mean near zero.Source: §3.3.4 · p. 8

3.4 Index

The study uses CCI30, a rules-based index tracking the 30 largest cryptocurrencies by market capitalization, and evaluates January 2019 daily close-price prediction and forecasting.

3.4 Index

  • CCI30 tracks the 30 largest cryptocurrencies by market capitalization to measure the blockchain sector’s overall growth and daily and long-term movement.Its starting value was set arbitrarily at 100 on January 1, 2015, and it launched on January 1, 2017.Source: §3.4 · p. 9
  • The study downloaded the index’s daily OHLCV data and targeted its daily close price for January 2019.The index data came from [34].Source: §3.4 · p. 9
  • For January close-price prediction, the study used gradient-boosted trees, ensemble learning, and a neural network; K-NN was used for forecasting.The prediction used training and testing datasets, while K-NN forecasting used only the training phase.Source: §3.4 · p. 9
  • Prediction excluded volume and market capital attributes, while forecasting used only Date and Close Price.Source: §3.4 · p. 9

3.5 Comparison with Previous Results

The section compares the paper’s model results with prior cryptocurrency studies, whose forecasting targets and data differ. The authors report favorable comparisons, while noting one higher accuracy result from a study with a different setup.

3.5 Comparison with Previous Results

  • Ensemble learning achieved accuracy 0.924, below LightGBM’s 0.952 in, which forecast price trends using six months of data rather than yearly data.The studies also differ in target: classified price trends, while this paper predicted and forecast prices for nine cryptocurrencies and CCI30.Source: §3.5 · p. 9
  • For Bitcoin, Dogecoin, and NEM, this paper reports gradient-boosted-trees RMSEs of 32.863, 0.000, and 0.001, versus LSTM RMSEs of 2.75 × 103, 19.2923, and 0.0499 in.The cited values refer to Bitcoin, Digital Cash, and Ripple, respectively.Source: §3.5 · p. 9
  • Against, the authors report maximum accuracy/RMSE pairs of 0.900/0.001 for gradient-boosted trees and 0.924/0.002 for ensemble learning, describing their maxima as much better.Source: §3.5 · p. 10

4 Conclusion

The paper presents four machine-learning models for predicting and forecasting close prices across nine cryptocurrency constituents and the cci30 index. It reports strong overall prediction performance and competitive results against prior work, but says K-NN forecasting was less effective amid noisy features and extreme volatility.

4 Conclusion

  • Four machine-learning models predict and forecast close prices for nine cryptocurrency constituents and the cci30 index.The conclusion names the models and assets as the paper’s scope.Source: §4 · p. 10
  • The models show good overall prediction performance and results that appear better and competitive with state-of-the-art work.The conclusion does not specify which model wins or provide numerical comparisons here.Source: §4 · p. 10
  • K-NN forecasting was less effective, which the authors attribute to noisy random features and extreme volatility.The authors contrast this forecasting result with the other models’ performance.Source: §4 · p. 10

S1: Performance vector of Gradient Boosted Trees model

The gradient-boosted-trees performance vector reports error, correlation, and trend-accuracy measures for cryptocurrency constituents. The listed results vary by constituent and metric.

S1: Performance vector of Gradient Boosted Trees model

S2: A sample of tree leaf obtained from gradient boosted trees model

The supplement provides a sample tree leaf from the gradient-boosted-trees model.

S2: A sample of tree leaf obtained from gradient boosted trees model

  • A sample leaf from the RapidMiner gradient-boosted-trees model is shown.The supplied passage identifies the illustration as a sample tree leaf.Source: Figure 2s · p. 29

S3: Performance vector of neural network model

The neural-network performance vectors report error, correlation, and prediction trend accuracy measures. The listed trend-accuracy values differ across the named constituents.

S3: Performance vector of neural network model

S4: Performance vector of ensemble learning method

The ensemble learner’s performance vector reports prediction metrics for multiple cryptocurrency constituents, with results varying across currencies.

S5: Performance vector of K-NN model

The K-NN performance vector reports constituent-level prediction metrics, while a separate figure presents a model comparison for cryptocurrency index 30.

Loading 1905.08444v1…