My trading strategy here is only based on BUY/SELL of the classifier. When system returns BUY,it buys and holds until opposite signal. But I think it can't be done on EURUSD 1M time frame due to high volatility and so much noise. Am I right?
The code I wrote for the model is:
Code:
import pandas as pd #importing Pandas library for creating our dataframe
import numpy as np #importing Numpy library , it's used for numerical and array usage in Python
from sklearn.model_selection import train_test_split
from sklearn.externals import joblib #joblib for dumping and loading our classifier
from sklearn.ensemble import RandomForestClassifier #importing RandomForest classifier
from sklearn.metrics import classification_report,roc_auc_score,accuracy_score
import talib #importing this library to calculate indicators
#Now we create a dataframe from our OHLC data and name it df
df = pd.read_csv('Data.csv')
#We name the columns
df.columns = ['DateTime','Open','High','Low','Close','Volume']
#We used these features only, and DateTime is not important
df = df[['Open','High','Low','Close','Volume']]
#At first we want to calculate Stochastic %K, I prefer to calculate it manually, so we define L14 column and H14 which are based on Lowest and Highest of past 14 days
df['L14'] = df['Low'].rolling(window=14).min()
df['H14'] = df['High'].rolling(window=14).max()
#Now we can calculate %K based on these above columns
df['%K'] = 100 * ((df['Close'] - df['L14']) / (df['H14'] - df['L14']))
#Now we used ta_lib library to calculate MACD,PROC,RSI, Williams R,OBV
#Price rate of Change
df['PROC'] = talib.ROCP(np.array(df.Close),1)
#RSI(14)
df['RSI'] = talib.RSI(np.array(df.Close))
#Williams R percentage
df['%R'] = talib.WILLR(np.array(df.High),np.array(df.Low),np.array(df.Close))
#MACD (12,26,9) NOTICE: WE NEED ONLY MAIN LINE,NOT SIGNAL LINE
df['MACD'] = talib.MACDFIX(np.array(df.Close))[0]
#On Balance Volume
df['OBV'] = talib.OBV(np.array(df.Close),np.array(df.Volume))
#Now we define our target variable, We want to predict next candlestick direction,so shift here is 1
df['Target'] = np.sign(df.Close.shift(1) - df.Close)
#Now we need to drop the columns which we have no interest in
df = df.drop(['Open','Low','Close','High','Volume','L14','H14'],axis=1)
#And also We drop all the NaN values
df = df.dropna()
#We create another dataframe which only has our features from the original dataframe and named it X_df
#And create another dataframe from the original dataframe which only has our target
X_df = df[['PROC','RSI','%R','MACD','OBV','%K']]
Y_df = df[['Target']]
#Y_val has values of dataframe in NdArray type
Y_val = Y_df.values
#Now we want to apply exponentially smoothing to our dataframe and get it's data as matrix, we named it dataset
#If we remove these two lines, we will get 50% accuracy!!
dataset = X_df.ewm(span=windowSize, min_periods=windowSize,).mean().as_matrix()
dataset = dataset[windowSize:]
#It creates a window with lookback of 10
def create_dataset(dataset, look_back=10):
dataX, dataY = [], []
for i in range(len(dataset)-look_back-1):
a = dataset[i:(i+look_back), 0:6]
dataX.append(a)
dataY.append(Y_val[i + look_back, 0])
return np.array(dataX), np.array(dataY)
#X,Y are the NdArray values which window lookback applied
X,Y = create_dataset(dataset)
#We reshape X to have 2 dimension to input it to our algorithm
X = X.reshape(len(X),-1)
#Making sure Y is integer
Y = Y.astype('int')
#Spliting the X,Y with test size of 0.2 , It splits the X,Y values to 20% for testing, 80% for training
trainX, testX, trainY, testY = train_test_split(X,Y,test_size=0.2)
#Defining classifier variable as one instance of RandomForestClassifer
classifier = RandomForestClassifier(n_jobs=-1,n_estimators=100,verbose=True,oob_score=True)
classifier.fit(trainX,trainY)
predicted = classifier.predict(testX)
print classifier.score(testX,testY)
print classification_report(testY,predicted)
print classifier.oob_score_
Result :
Accuracy Score : 0.878622197922362
precision recall f1-score support
-1 0.87 0.95 0.91 14395
0 0.84 0.04 0.08 2239
1 0.88 0.94 0.91 14459
avg / total 0.88 0.88 0.85 31093
OOB Score: 0.8701154600713987
With some optimizing and minimizing the window and look back we can get accuracy as high as 94%
If we removing Doji from our dataframe we will have:
Accuracy Score: 0.9407559087911707
precision recall f1-score support
-1 0.94 0.95 0.94 14411
1 0.94 0.94 0.94 14402
avg / total 0.94 0.94 0.94 28813
OOB Score: : 0.9349929283551267
ROC score :0.9407544394884746
Am I doing right? When I tested it on Demo account,it didn't work as expected. Maybe EURUSD on 1M is so noisy that it can't work as expected. What is the problem here?
