Cryptocurrency Buy Signal Prediction with Deep Neural Network – Q Learning

This simple TensorFlow app tries to learn is the best time to buy Crypto Asset. For this goal use the Bitcoin historical data (5-minute chart) and validate the trained model with the Ethereum historical data. The basic concept is to create a simple model and provide all the information for it. The model's input parameters are the candle’s data and technical indicators. Moreover, the model gets separate signals when the indicators give a buy signal.

Trading logic: The model learns what are the good conditions to make a long order. Using the ATR indicator to determine the SL and TP values. The default SL value is 2 times the ATR value. The TP values depend on the market's conditions. The model can return three different types of buy signals.

Buy1: Risk Reward ratio is 1:1.5 when the TP is 1.5 times the SL value. (3 times ATR)

Buy2: Risk Reward ratio is 1:2 when the TP is 2 times the SL risk. (4 times ATR)

Buy3: Risk Reward ratio is 1:3 when the TP is 3 times the SL values (6 times ATR)

Used Indicators:

• EMA – 200 period of Exponential Moving Average, buy signal when the price is above the EMA line.
• MACD – MACD line and Signal line, buy signal MACD cross (all MACD cross not filtered out the above 0 crosses )
• CMF – Chaikin Money Flow indicator, buy signal when the indicator is above 0
• ATR – Average True Rage – determinate the Stop Loss level and Take Profit level.
• RSI – Relative Strength Index, buy signal when the indicator shows oversold (below 30), extra buy signal when RSI left the 30 level to up.
• ADX – Trend Strength Indicator (14 candles), buy signal when the indicator is above 50
• ADX – Trend Strength Indicator (288 candles), buy signal when the indicator is above 50
• SMA – Simple Moving Average 1440 periods (5 days) buy signal when the price is above the SMA line
• STOCHASTIC K line and STOCHASTIC D line buy signal when stochastic below 20 extra buy signal when K line cross D line below 20
• Bollinger Bands Upper and lower brand with default settings

Hyperparameters

• ACTIVATION: default ‘Sigmoid’ activation function for the keras model.
• GAMMA: default 0.99 set how important is the current reward
• LR: default 0.1 initial learning rate, shows how important the current values.
• EPSILON_DEC: default 1e-5 epsilon decay value
• EPSILON_START default 1 random step /predicted step rate start value
• EPSILON_END default 0.1 random step /predicted step rate minimum value

Parameters:

• N_TRADE: number of trades. The app stops after this amount of trade
• RENDER_MODE: 'compueter' or 'human' show or hide the chart data. Human mode not implemented *
• CANDLES: default 72 sets up how many candles are the input parameter.

Monitor learning progress using the TensorBoard

Run the following command from the source folder

tensorboard --logdir './logs/train' 

DISCLAIMER

I am not a financial advisor. This code is for learning purposes only. Trading with crypto assets is extremely risky. The value of your investment can go down as well as up, and you may get back less than you invest.