UNSURE - A machine learning approach to cryptocurrency trading

Abstract

Although cryptocurrency trading can be highly profitable, it carries significant risks due to extreme price fluctuations and high degree of market noise. To increase profits and minimize risks, traders typically use various forecasting methods, such as technical analysis and Machine Learning (ML), but develo** effective trading strategies in noisy markets still remains a challenging task. Recently, Deep Reinforcement Learning (DRL) agents have achieved high performance on challenging tasks, including algorithmic trading, however it requires significant amount of time and high-quality data to train effectively. Additionally, DRL agents lack explainability, making them a less popular option for traders. The purpose of this paper is to address these challenges by proposing a reliable trading framework. Our framework, named UNSURE, generates high-quality features from candlestick data using technical analysis along with a novel parameterization method, and then exploits high price fluctuations by combining three ML components: A) Unsupervised component, which further improves feature quality by clustering market data; B) DRL component, which is responsible for training agents that open Buy or Short positions; C) Supervised component, which estimates price fluctuations in order to open and close positions efficiently, while reducing trading uncertainty. We demonstrate the effectiveness of this approach on nine cryptocurrency markets using several risk-adjusted performance metrics.

Appendices

Appendix A technical analysis

1.1 A.1 Exponential moving average - EMA

EMA is a moving average type of indicator, which is given by (A1). EMA is used to smooth price signals, in order to remove market noise.

$$\begin{aligned} EMA(n) = Price(n)*k + EMA(n - 1)*(1 - k) \end{aligned}$$

(A1)

where k is the smoothing constant and is defined as $k=\frac{2}{N + 1}$. N parameter defines the "Look-Back" sliding window and is used in many technical indicators.

1.2 A.2 Double-exponential moving average - DEMA

DEMA is an extension of EMA, given by (A2), which attempts to remove price lag that is caused by EMA. The default value of N parameter is 14.

$$\begin{aligned} DEMA_{N} = 2EMA_{N} - EMA \, of \, EMA_{N} \end{aligned}$$

(A2)

1.3 A.3 Moving average convergence/divergence - MACD

MACD is also a trend type of indicator, which shows the relationship between two moving averages (usually EMAs), one with a short period and one with a longer period. The mathematical formula of MACD is provided by (A3).

$$\begin{aligned} MACD = EMA_{N_{1}} - EMA_{N_{2}} \end{aligned}$$

(A3)

with $N_{2}>> N_{1}$. The default values are $N_{1} = 12$ and $N_{2} = 26$

1.4 A.4 Aroon UP/DOWN

Aroon, which is described by (A4) and (A5), identifies trend changes and estimate the strength of an upcoming trend.

$$\begin{aligned} \frac{25 - N_{H}}{25}*100 \end{aligned}$$

(A4)

$$\begin{aligned} \frac{25 - N_{L}}{25}*100 \end{aligned}$$

(A5)

with $N_{H}$ and $N_{L}$ being the timesteps between a new High price or Low price respectively.

1.5 A.5 Commodity channel index - CCI

CCI is also trend indicator, which estimates price trends, as well as the direction and strength of a trend. CCI is presented by (A6).

$$\begin{aligned} CCI = \frac{TP(n) - EMA_{N}(TP)}{MD(TP)}*0.015 \end{aligned}$$

(A6)

where, TP the typical price defined as $TP(n) = \frac{High(n) + Low(n) + Close(n)}{3}$ and MD(TP) the Mean Deviation of Typical Price. Typically, $N=20$.

1.6 A.6 Average directional index - ADX

ADX is a well know trend type of indicator that measures the strength of a trend and is given by (A7)

$$\begin{aligned} ADX = MA * \frac{PDI - NDI}{PDI + NDI}*100 \end{aligned}$$

(A7)

where MA is a Moving Average and PDI, NDI are Positive Directional Indicator and Negative Directional Indicator respectively.

1.7 A.7 Stochastic oscillator - STOCH

STOCH is a momentum indicator, given by (A8), which is usually used in large trading ranges or to capture slow moving trends.

$$\begin{aligned} \frac{Close(n) - L_{N}}{H_{N} - L_{N}} \end{aligned}$$

(A8)

where $L_{N}$ and $H_{N}$ are the minimum past value Low value and maximum past High value respectively, within a window of N prices. The default value of N is 14.

1.8 A.8 Relative strength index - RSI

RSI is a well-studied momentum indicator, which attempts to identify over-bought or over-sold securities. The mathematical formula of RSI is described by (A9)

$$\begin{aligned} RSI = 100 - \frac{100}{1 + RS} \end{aligned}$$

(A9)

where RS is the ratio of Average Price Increases to Average Price Drops within a window of N timesteps. The default period is 14.

1.9 A.9 On-balance volume - OBV

OBV is a volume type of indicator that measures the volume flow. Its formula is provided by (A10).

$$\begin{aligned} \begin{gathered} OBV(n+1) = OBV(n) + \\ \left\{ \begin{array}{ll} Volume(n) &{} Close(n + 1) < Close(n) \\ 0 &{} Close(n + 1) = Close(n) \\ -Volume(n) &{} Close(n + 1) > Close(n) \end{array} \right\} \end{gathered} \end{aligned}$$

(A10)

1.10 A.10 Bolliger bands - BBANDS

BBANDS is a volatility indicator, which is provided by (A11) and (A12) and measures the standard deviation of the simple moving average.

$$\begin{aligned} BBAND_{UP} = Mean(TP) + 2*Std(TP) \end{aligned}$$

(A11)

$$\begin{aligned} BBAND_{DOWN} = Mean(TP) - 2*Std(TP) \end{aligned}$$

(A12)

where Mean(TP) is the average typical price and Std(TP) is the standard deviation of typical price within a look back window N. The default window of Bollinger-Bands is $N=20$.

1.11 A.11 Volume-weighted average price - VWAP

VWAP is a volume-weighted moving average, which is given by (A13) and defined as the average price of an asset weighted by the total trading volume over a period of N timesteps. The default period is usually $N=14$.

$$\begin{aligned} VWAP(n) = \frac{\sum _{n=0}^{N - 1} TP(N - i)*Volume(N - i)}{\sum _{n=0}^{N - 1} Volume(N - i)} \end{aligned}$$

(A13)

1.12 A.12 Accumulation/distribution line - ADL

ADL is another volume-based indicator, which attempts to measure the underlying supply and demand (bids and asks). The mathematical formula of ADL indicator is described by (A14).

$$\begin{aligned} ADL(n) = ADL(n-1) + \frac{(Close - Low)-(High - Close)}{High - Low} \end{aligned}$$

(A14)

Appendix B Sensitivity analysis

Table 10 Analysis of Budget and Transaction Volume parameters in the evaluation datasaet of Bitcoin market

Fig. 12

Average Maximum Volatility (left) and the respective TCN’s Mean Absolute Log Error (right) per Horizon in Bitcoin Market. The orange point represents the selected value of our framework

Table 11 UNSURE’s perforance in terms of PNL and ADD metrics in respect to Virtual Fees parameter in the evaluation dataset of Bitcoin market

Table 12 Analysis of TCN’s performance in respect to Pinball Loss Quantile ($\tau$) parameter in the evaluation dataset of Bitcoin market