結合多層1D CNN和LSTM預測股價走勢
1D CNN具有短期特徵擷取與短期記憶功能,而LSTM則具有較長期的記憶功能,我們結合1D CNN和LSTM的優點來預測股價走勢,並與之前的實作案例多層LSTM模型比較 訓練資料與之前多層LSTM大致相同,唯一不能的地方是訓練資料的input為三維度向量(樣本,股價,ConvNet output)
import os
import pandas as pd
import matplotlib.pyplot as plt
import sqlite3
import numpy as np
from numpy import newaxis
import pandas as pd
from subprocess import check_output
from keras import layers
from keras.layers.core import Activation, Dropout, Flatten
from keras.layers import Dense
from keras.layers.recurrent import LSTM
from keras.models import Sequential
from sklearn.model_selection import train_test_split
import time
from sklearn.preprocessing import MinMaxScalerv
import matplotlib.pyplot as plt
import os
import pandas as pd
import matplotlib.pyplot as plt
import sqlite3
import numpy as np
from numpy import newaxis
import pandas as pd
from subprocess import check_output
from keras import layers
from keras.layers.core import Activation, Dropout, Flatten
from keras.layers import Dense
from keras.layers.recurrent import LSTM
from keras.models import Sequential
from sklearn.model_selection import train_test_split
import time
from sklearn.preprocessing import MinMaxScalerv
import matplotlib.pyplot as plt
載入資料庫
os.getcwd() #取得目前路徑
os.chdir('/Users/laihunglai/Dropbox/Tensorflow/stock')
os.getcwd()
db=sqlite3.connect('stock_by_index')
stock=pd.read_sql(con=db,sql='SELECT * FROM "0050"')
stock_price=stock['收盤價'].values.astype('float32')[:,np.newaxis]
scaler = MinMaxScaler(feature_range=(0, 1))
stock_price = scaler.fit_transform(stock_price)
plt.plot(stock_price)
plt.show()
train_size = int(len(stock_price) * 0.80)
test_size = len(stock_price) - train_size
train, test = stock_price[0:train_size,:], stock_price[train_size:len(stock_price),:]
print(len(train), len(test))
976 245
準備訓練資料的函數: 輸入為股價,輸出為dataX:m筆look_back天數的股價shape=(m,look_back)與dataY:m筆look_back後一天的股價shape=(m,1)
# convert an array of values into a dataset matrix
def create_dataset(dataset, look_back=1):
dataX, dataY = [], []
for i in np.arange(len(dataset)-look_back-1):
a = dataset[i:(i+look_back), 0]
dataX.append(a)
b=dataset[(i+look_back), 0]
dataY.append(b)
return np.array(dataX), np.array(dataY)
#將訓練和測試股價帶入上面的函數得到(trainX, trainY,),(testX, testY)
look_back = 30
trainX, trainY = create_dataset(train, look_back)
testX, testY = create_dataset(test, look_back)
trainX=trainX[:,:,np.newaxis]
testX=testX[:,:,np.newaxis]
trainX.shape
(945, 30, 1)
#建立多層LSTM,最後用密集層輸出一維度的預測值
model = Sequential()
model.add(layers.Conv1D(32,5,activation='relu',padding='same',input_shape=(None,trainX[0].shape[-1])))
model.add(layers.MaxPool1D(3))
model.add(layers.Conv1D(32,5,activation='relu',padding='same'))
model.add(layers.MaxPool1D(3))
model.add(LSTM(32,return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(32,return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(32,return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(32,return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(32,return_sequences=False))
model.add(Dropout(0.2))
model.add(Dense(output_dim=1))
model.add(Activation('linear'))
start = time.time()
model.compile(loss='mse', optimizer='rmsprop', metrics=['mae'])
print ('compilation time : ', time.time() - start)
model.summary()
________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv1d_1 (Conv1D) (None, None, 32) 192
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, None, 32) 0
_________________________________________________________________
conv1d_2 (Conv1D) (None, None, 32) 5152
_________________________________________________________________
max_pooling1d_2 (MaxPooling1 (None, None, 32) 0
_________________________________________________________________
lstm_1 (LSTM) (None, None, 32) 8320
_________________________________________________________________
dropout_1 (Dropout) (None, None, 32) 0
_________________________________________________________________
lstm_2 (LSTM) (None, None, 32) 8320
_________________________________________________________________
dropout_2 (Dropout) (None, None, 32) 0
_________________________________________________________________
lstm_3 (LSTM) (None, None, 32) 8320
_________________________________________________________________
dropout_3 (Dropout) (None, None, 32) 0
_________________________________________________________________
lstm_4 (LSTM) (None, None, 32) 8320
_________________________________________________________________
dropout_4 (Dropout) (None, None, 32) 0
_________________________________________________________________
lstm_5 (LSTM) (None, 32) 8320
_________________________________________________________________
dropout_5 (Dropout) (None, 32) 0
_________________________________________________________________
dense_1 (Dense) (None, 1) 33
_________________________________________________________________
activation_1 (Activation) (None, 1) 0
=================================================================
Total params: 46,977
Trainable params: 46,977
Non-trainable params: 0
_________________________________________________________________
train_history=model.fit(trainX,trainY,batch_size=128,nb_epoch=20,validation_split=0.2)
Train on 756 samples, validate on 189 samples
Epoch 1/20
756/756 [==============================] - 10s 13ms/step - loss: 0.1512 - mean_absolute_error: 0.3462 - val_loss: 0.4892 - val_mean_absolute_error: 0.6971
Epoch 2/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0616 - mean_absolute_error: 0.1950 - val_loss: 0.0910 - val_mean_absolute_error: 0.2958
Epoch 3/20
756/756 [==============================] - 2s 3ms/step - loss: 0.0205 - mean_absolute_error: 0.1135 - val_loss: 0.0843 - val_mean_absolute_error: 0.2840
Epoch 4/20
756/756 [==============================] - 2s 2ms/step - loss: 0.0171 - mean_absolute_error: 0.1028 - val_loss: 0.0177 - val_mean_absolute_error: 0.1170
Epoch 5/20
756/756 [==============================] - 2s 2ms/step - loss: 0.0126 - mean_absolute_error: 0.0892 - val_loss: 0.0354 - val_mean_absolute_error: 0.1767
Epoch 6/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0144 - mean_absolute_error: 0.0936 - val_loss: 0.0176 - val_mean_absolute_error: 0.1152
Epoch 7/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0117 - mean_absolute_error: 0.0862 - val_loss: 0.0052 - val_mean_absolute_error: 0.0570
Epoch 8/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0113 - mean_absolute_error: 0.0823 - val_loss: 0.0052 - val_mean_absolute_error: 0.0585
Epoch 9/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0119 - mean_absolute_error: 0.0869 - val_loss: 0.0332 - val_mean_absolute_error: 0.1685
Epoch 10/20
756/756 [==============================] - 2s 3ms/step - loss: 0.0115 - mean_absolute_error: 0.0848 - val_loss: 0.0070 - val_mean_absolute_error: 0.0704
Epoch 11/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0104 - mean_absolute_error: 0.0801 - val_loss: 0.0062 - val_mean_absolute_error: 0.0664
Epoch 12/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0088 - mean_absolute_error: 0.0739 - val_loss: 0.0317 - val_mean_absolute_error: 0.1616
Epoch 13/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0115 - mean_absolute_error: 0.0858 - val_loss: 0.0099 - val_mean_absolute_error: 0.0791
Epoch 14/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0083 - mean_absolute_error: 0.0726 - val_loss: 0.0411 - val_mean_absolute_error: 0.1871
Epoch 15/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0134 - mean_absolute_error: 0.0930 - val_loss: 0.0088 - val_mean_absolute_error: 0.0740
Epoch 16/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0088 - mean_absolute_error: 0.0741 - val_loss: 0.0073 - val_mean_absolute_error: 0.0722
Epoch 17/20
756/756 [==============================] - 2s 3ms/step - loss: 0.0095 - mean_absolute_error: 0.0778 - val_loss: 0.0056 - val_mean_absolute_error: 0.0620
Epoch 18/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0097 - mean_absolute_error: 0.0781 - val_loss: 0.0059 - val_mean_absolute_error: 0.0641
Epoch 19/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0078 - mean_absolute_error: 0.0696 - val_loss: 0.0066 - val_mean_absolute_error: 0.0630
Epoch 20/20
756/756 [==============================] - 1s 2ms/step - loss: 0.0113 - mean_absolute_error: 0.0835 - val_loss: 0.0065 - val_mean_absolute_error: 0.0625
import matplotlib.pyplot as plt
def show_train_history(train_history, train, validation):
plt.plot(train_history.history[train])
plt.plot(train_history.history[validation])
plt.title('Train history')
plt.ylabel(train)
plt.yscale('log')
plt.xlabel('Epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.show()
show_train_history(train_history,'loss','val_loss')
show_train_history(train_history,'mean_absolute_error','val_mean_absolute_error')
def predict_sequences_multiple(model, firstValue,length):
prediction_seqs = []
curr_frame = firstValue
for i in range(length):
a=model.predict(curr_frame[np.newaxis,:])[0,0]
curr_frame=np.insert(curr_frame,look_back,a)[-look_back:][:,newaxis]
prediction_seqs.append(a)
# print(curr_frame.reshape(1,30))
return prediction_seqs
# print(curr_frame.reshape(curr_frame.shape[1],curr_frame.shape[0]))
def inverse_transform(testY,predictions):
predictions_origins=scaler.inverse_transform(np.array(predictions).reshape(-1, 1))
answer=scaler.inverse_transform(testY.reshape(-1,1))
return answer,predictions_origins
def stock_pred_plot(answer,predictions_origins):
error=(predictions_origins-answer)/answer*100
fig,axis=plt.subplots(2,1,figsize=(5,4))
axis[0].plot(predictions_origins,label='predict')
axis[0].plot(answer,label='real')
axis[0].set_ylabel('price')
axis[0].legend()
axis[1].set_xlabel('Day')
axis[1].set_ylabel('error(%)')
axis[1].plot(error,label='error')
axis[1].legend()
plt.show()
def error_history(answer,predictions_origins):
result=(predictions_origins-answer)/answer*100
return result[:,0]
predict_length=30
for i in np.arange(0,100,5):
predictions = predict_sequences_multiple(model, testX[i], predict_length)
answer=testY[i:(i+predict_length)]
ans,pred=inverse_transform(answer,predictions)
stock_pred_plot(ans,pred)
#計算誤差值
error_history_all=[]
for i in np.arange(0,100,5):
predictions = predict_sequences_multiple(model, testX[i], predict_length)
answer=testY[i:(i+predict_length)]
ans,pred=inverse_transform(answer,predictions)
error_history_all.append(error_history(ans,pred))
all_error=np.array(error_history_all)
mean=all_error.mean(axis=0)
std=all_error.std(axis=0)
df=pd.DataFrame(all_error)
#畫出測試誤差
fig,axis=plt.subplots(2,1, figsize=(10,5))
df.T.plot(ax=axis[0])
axis[0].set_ylabel('Error(%)')
import matplotlib.pyplot as plt
x=list(range(len(df.T)))
axis[1].errorbar(x, mean, std, linestyle='None', marker='^')
axis[1].set_xlabel('Day')
axis[1].set_ylabel('Error(%)')
Text(0,0.5,'Error(%)')
