【深度学习量化投资】RNNs在股票价格预测的应用基于Keras

前言

RNN和LSTMs在时态数据上表现特别好,这就是为什么他们在语音识别上是有效的。我们通过前25天的开高收低价格,去预测下一时刻的前收盘价。每个时间序列通过一个高斯分布和2层LSTM模型训练数据。文章分为两个版块,外汇价格预测和每日盘中价格预测(30分钟、15分钟、5分钟,等等)。源代码请在文末获取!

外汇预测(用英语描述)

a. Daily Data is pulled from Yahoo’s Data Reader

b. Only the training set is preprocessed because we create a separate test set later on

c. “model_forex” is the model for to build and train.

d. Create separate daily test set by specifying dates which start after your training set ends.

e. You can see “model_forex” is plugged in here for running the prediction

predicted_st = predict_standard(X_test_stock,y_test_stock, model_forex)

盘中预测(用英语描述)

a. Intraday Data is pulled from Google’s API. The second argument is the time in seconds (900 secs = 15 mins) and the third argument it the number of days, the max backtrack day for Googles API is 15 days I believe.

df = get_google_data(INTRA_DAY_TICKER, 900, 150)

b. Preprocess the full set of data and train test split it with “train_test_split_intra”

c. “model_intra” is the model for to build and train.

d. You can see “model_intra” is plugged in here for running the prediction

predicted_intra = predict_intra(X_test_intra,y_test_intra, model_intra)

代码展示

SITE = "http://en.wikipedia.org/wiki/List_of_S%26P_500_companies"def scrape_list(site):
    hdr = {'User-Agent': 'Mozilla/5.0'}
    req = urllib2.Request(site, headers=hdr)
    page = urllib2.urlopen(req)
    soup = BeautifulSoup(page)

    table = soup.find('table', {'class': 'wikitable sortable'})
    sector_tickers = dict()    for row in table.findAll('tr'):
        col = row.findAll('td')        if len(col) > 0:
            sector = str(col[3].string.strip()).lower().replace(' ', '_')
            ticker = str(col[0].string.strip())            if sector not in sector_tickers:
                sector_tickers[sector] = list()
            sector_tickers[sector].append(ticker)    return sector_tickers
sector_tickers = scrape_list(SITE)##Help functions to normalize and denormalize values(省略)
# Sequence Length, or # of days of tradingSEQ_LENGTH = 25# Number of units in the two hidden (LSTM) layersN_HIDDEN = 256#Number of attributes used for each trading daynum_attr = 4#Out of those attribute how many are indicatorsnum_indicators = 0#Variable to help define how far you want your y to reachREWARD_LAG = 1#How many dats ahead do you want to predictLOOK_AHEAD = 5#Window StrideSTRIDE = 1
def _load_data(data, n_prev = SEQ_LENGTH):  
    docX, docY = [], []    for i in range(len(data)-n_prev):
        x,y = norm(data.iloc[i:i+n_prev,:num_attr].as_matrix(),data.iloc[i+n_prev-1,num_attr:].as_matrix())
        docX.append(x)
        docY.append(y)
    alsX = np.array(docX)
    alsY = np.array(docY)    return alsX, alsYdef _load_data_test(data, n_prev = SEQ_LENGTH):  
    docX, docY = [], []
    num_sequences = (len(data)-n_prev+1)/STRIDE    for i in range(num_sequences):
        i = i*STRIDE
        x = (data.iloc[i:i+n_prev,:num_attr].as_matrix())
        y = (data.iloc[i+n_prev-1,num_attr:].as_matrix())        #x,y = norm(data.iloc[i:i+n_prev,:num_attr].as_matrix(),data.iloc[i+n_prev-1,num_attr:].as_matrix())
        docX.append(x)
        docY.append(y)
    alsX = np.array(docX)
    alsY = np.array(docY)    return alsX, alsYdef _load_data_norm(data, n_prev = SEQ_LENGTH):  
    docX, docY = [], []    for i in range(len(data)-n_prev):
        x = np.array((data.iloc[i:i+n_prev,:num_attr].as_matrix()))
        y = np.array((data.iloc[i+n_prev-1,num_attr:].as_matrix()))
(省略)

外汇数据

##Dataset on just single ticker to test performancesdf = data.DataReader('EUR=X', 'yahoo', datetime(2010,8,1), datetime(2014,8,1))# df['RSI'] = ta.RSI(df.Close.values,timeperiod=14)# _,_, macdhist = ta.MACD(df.Close.values, fastperiod=12, slowperiod=26, signalperiod=9)# df['MACDHist'] = macdhist##Add the predicted coloumn Y, as the last coloumn can be defined however you think is a good representation of a good decision ##Clean the rest of the Data Framey = []for i in range(0,len(df)):    if i >= (len(df)- STRIDE):
        y.append(None)    else:        if (REWARD_LAG > 1):
            val = 0
            for n in range(REWARD_LAG):
                val = val + df['Close'][i+n+1]
            val = val / float(REWARD_LAG)
            y.append(val)        else:
            y.append(df['Close'][i+REWARD_LAG])


df['Y_Values'] =np.asarray(y)
df = df.dropna()#print (df)sliced_df = df.drop(['Adj Close','Volume'] ,axis=1)#print (sliced_df)#(X_train, y_train), (X_test, y_test) = train_test_split(sliced_df)(X_train, y_train) = train_test_split(sliced_df)
print(X_train[0],y_train[0])print (X_train.shape,y_train.shape)
(array([[-0.76244909, -0.75153814, -1.36800657, -1.28695383],
       [-1.28305706, -1.17005084, -1.66649887, -1.50673145],

(省略)

盘中数据

def get_google_data(symbol, period, window):
    url_root = 'http://www.google.com/finance/getprices?i='
    url_root += str(period) + '&p=' + str(window)
    url_root += 'd&f=d,o,h,l,c,v&df=cpct&q=' + symbol
    print(url_root)
    response = urllib2.urlopen(url_root)
    data = response.read().split('\n')    #actual data starts at index = 7
    #first line contains full timestamp,
    #every other line is offset of period from timestamp
    parsed_data = []
    anchor_stamp = ''
    end = len(data)    for i in range(7, end):
        cdata = data[i].split(',')        if 'a' in cdata[0]:            #first one record anchor timestamp
            anchor_stamp = cdata[0].replace('a', '')
            cts = int(anchor_stamp)        else:            try:
                coffset = int(cdata[0])
                cts = int(anchor_stamp) + (coffset * period)
                parsed_data.append((dt.datetime.fromtimestamp(float(cts)), float(cdata[1]), float(cdata[2]), float(cdata[3]), float(cdata[4]), float(cdata[5])))            except:                pass # for time zone offsets thrown into data
    df = pd.DataFrame(parsed_data)
    df.columns = ['ts', 'Open', 'High', 'Low', 'Close', 'Volume']
    df.index = df.ts    del df['ts']    return df

盘中创建单独的数据集

df = get_google_data('AAPL', 900, 150)#print(df)plt.plot(df['Close'].values[:])
y = []for i in range(0,len(df)):    if i >= (len(df)- REWARD_LAG):
        y.append(None)    else:        if (REWARD_LAG > 1):
            val = 0
            for n in range(REWARD_LAG):
                val = val + df['Close'][i+n+1]
            val = val / float(REWARD_LAG)
            y.append(val)
            print('here')        else:
            y.append(df['Close'][i+REWARD_LAG])


df['Y_Values'] =np.asarray(y)
df = df.dropna()
sliced_df = df.drop(['Volume'] ,axis=1)#print(sliced_df)(X_train, y_train), (X_test, y_test) = train_test_split_intra(sliced_df)#print(X_train[0],y_train[0])print(len(X_train),len(X_test))#print(X_test[0],y_test[0])
(1168, 108)

构建网络结构

model_intra = Sequential() 

model_intra.add(LSTM(N_HIDDEN, return_sequences=True, activation='tanh', input_shape=(SEQ_LENGTH, num_attr)))#model_intra.add(LSTM(N_HIDDEN, return_sequences=True, activation='tanh'))model_intra.add(LSTM(N_HIDDEN, return_sequences=False, activation='tanh'))

model_intra.add(Dense(1,activation='linear'))
model_intra.compile(loss="mean_squared_error", optimizer='adam')
model_intra_full = Sequential() 

model_intra_full.add(LSTM(N_HIDDEN, return_sequences=True, activation='tanh', input_shape=(SEQ_LENGTH, num_attr)))#model_intra_full.add(LSTM(N_HIDDEN, return_sequences=True, activation='tanh'))model_intra_full.add(LSTM(N_HIDDEN, return_sequences=False, activation='tanh'))

model_intra_full.add(Dense(1,activation='linear'))
model_intra_full.compile(loss="mean_squared_error", optimizer='adam')

model_forex = Sequential() 

model_forex.add(LSTM(N_HIDDEN, return_sequences=True, activation='tanh', input_shape=(SEQ_LENGTH, num_attr)))#model_forex.add(LSTM(N_HIDDEN, return_sequences=True, activation='tanh'))model_forex.add(LSTM(N_HIDDEN, return_sequences=False, activation='tanh'))

model_forex.add(Dense(1,activation='linear'))
model_forex.compile(loss="mean_squared_error", optimizer

符合模型的模型和参数

print(X_train.shape)
print(y_train.shape)
(1018, 25, 4)
(1018, 1)
model_intra.fit(X_train, y_train, batch_size=50, nb_epoch=
Train on 1156 samples, validate on 12 samples
Epoch 1/150
1156/1156 [==============================] - 1s - loss: 1.9575 - val_loss: 0.5494
Epoch 2/150
1156/1156 [==============================] - 1s - loss: 1.4731 - val_loss: 0.4006

(省略)

辅助绩效评估功能

#Function to normalize the test input then denormalize the result. Calculate the rmse of the predicted values on the test setdef predict(X_test,y_test, myModel):
    predicted = []    for example in X_test:
        x = copy.copy(example)        #print (x)
        x_norm, mn, mx = normalize(x)
        toPred = []
        toPred.append(x_norm)
        add = np.array(toPred)        #Predict for the standard model
        predict_standard = myModel.predict(add)
        pred_st = copy.copy(predict_standard)
        y_real_st = deNormalizeY(pred_st,mn,mx)
        predicted.append(y_real_st[0])        #Predict for the bidirectional model#         predict_bidirectional = myModel.predict([add,add])#         pred_bi = copy.copy(predict_bidirectional)#         y_real_bi = deNormalizeY(pred_bi,mn,mx)#         predicted.append(y_real_bi[0])(省略)df_test = data.DataReader('EUR=X', 'yahoo', datetime(2014,8,1), datetime(2015,8,1))# df_test['RSI'] = ta.RSI(df_test.Close.values,timeperiod=14)# _,_, macdhist = ta.MACD(df_test.Close.values, fastperiod=12, slowperiod=26, signalperiod=9)# df_test['MACDHist'] = macdhisty = []for i in range(0,len(df_test)):    if i >= (len(df_test)- STRIDE):
        y.append(None)    else:        if (REWARD_LAG > 1):
            val = 0
            for n in range(REWARD_LAG):
                val = val + df_test['Close'][i+n+1]
            val = val / float(REWARD_LAG)
            y.append(val)        else:
            y.append(df_test['Close'][i+REWARD_LAG])

(省略)
MAE for LSTM is: [0.0035823152701196983]
MAE for doing nothing is: [0.0045693478326778786]
RMSE for LSTM is: [0.0050684837061917686]
RMSE for doing nothing is: [0.0061416562709802761]
Net profit for 0.0 threshhold is 245.261025777 making 234 trades
Net profit for 0.001 threshhold is 242.673572498 making 201 trades
(省略)

盘中交易评价和结果

def predict_intra(X_test, y_test, myModel):
    print(len(X_test))
    predicted = []    for example in X_test:        #Transform the training example into gaussing distribution
        x_norm, mean, std = normDist(np.array(example))        #Add examples to array to predict
        toPred = []
        toPred.append(x_norm)
        add = np.array(toPred)        #Predict these examples
        predict_standard = myModel.predict(add)
        pred = copy.copy(predict_standard)
        y_real = deNormDist(pred,mean,std)
        predicted.append(y_real[0])    return predicted

predicted_intra = predict_intra(X_test,y_test, model_intra)
plt.figure(figsize=(20,20))
plt.plot(y_test) 
plt.plot(predicted_intra)
plt.show()

MAE and RMSE 评估

sum_error = 0sum_error_donothing = 0for i in range(len(predicted_intra)):    if i>0:
        sum_error = sum_error + abs(predicted_intra[i] - y_test[i])
        sum_error_donothing = sum_error_donothing + abs(predicted_intra[i] - y_test[i-1])
MAE_lstm = sum_error/len(predicted_intra)
MAE_donothing = sum_error_donothing/len(predicted_intra)
print("MAE for LSTM is: " + str(MAE_lstm))
print("MAE for doing nothing is: " + str(MAE_donothing))
MAE for LSTM is: [0.091961468484759237]
MAE for doing nothing is: [0.16699238882416201]
sum_error = 0sum_error_donothing = 0for i in range(len(predicted_intra)):    if i>0:
        sum_error = sum_error + (predicted_intra[i] - y_test[i])**2
        sum_error_donothing = sum_error_donothing + (predicted_intra[i] - y_test[i-1])**2RMSE_lstm = (sum_error/len(predicted_intra))**(1.0/2.0)
RMSE_donothing = (sum_error_donothing/len(predicted_intra))**(1.0/2.0)
print("RMSE for LSTM is: " + str(RMSE_lstm))
print("RMSE for doing nothing is: " + str(RMSE_dono
RMSE for LSTM is: [0.15719269057322682]
RMSE for doing nothing is: [0.23207816758496383]

Policy的功能评价

net_profits = []
protits_per_trade = []for i in range(50):
    THRESH = i/10000.0
    LOT_SIZE = 100
    net_profit = 0
    num_trades = 0
    for i in range(len(predicted_intra)):        if i>1:
            predicted_change = ((predicted_intra[i] / y_test[i-1]) - 1)            #print(predicted_change)
            actual_change = (predicted_intra[i] -  y_test[i])*LOT_SIZE            if predicted_change >= THRESH:                #print("Buy")
                net_profit = net_profit + actual_change 
                num_trades = num_trades + 1(省略)
(array([327.67074597699519], dtype=object), 106)
(array([322.81673063817777], dtype=object), 103)
plt.plot(net_profits)
plt.show()
plt.plot(protits_per_trade)
plt.show()

其他

buyTotal = 0sellTotal = 0correct = 0sellCorrect = 0buyCorrect = 0for i in range(len(predicted_st)):
    realAnswer = y_test_stock[i][0][0]    if predicted_st[i][1] > predicted_st[i][0]:
        predicted = 0 #Buy
    else:
        predicted = 1 #Sell

    if realAnswer == 0:        ##This is where the actual answer is Buy:Up:[0,1]:0
        buyTotal = buyTotal + 1
        if predicted == realAnswer:
            buyCorrect = buyCorrect + 1
            correct = correct + 1(省略)
(349, 730, 0.4780821917808219)
(210, 382, 0.5497382198952879)
(139, 348, 0.3994252873563218)
0.523287671233
0.476712328767
MMM
AYI
ALK
ALLE(省略)

创造基线RMSE

totalCorrect = 0total = 0for stock in testing_dataframes[:50]:

    X_test_stock, y_test_stock = _load_data_test(stock[1])
    predicted_st = predict_standard(X_test_stock,y_test_stock, model)

    buyTotal = 0
    sellTotal = 0
    correct = 0
    sellCorrect = 0
    buyCorrect = 0(省略)
#Count the number of positive and the number of negative calls you got righttotalCorrect = 0total = 0buyTotal = 0sellTotal = 0correct = 0sellCorrect = 0buyCorrect = 0for i in range(len(predicted_st)):
    realAnswer = y_test_stock[i][0][0]    if predicted_st[i][1] > predicted_st[i][0]:
        predicted = 0 #Buy(省略)
(104, 235, 0.4425531914893617)
(104, 104, 1.0)
(0, 131, 0.0)
0.442553191489
0.557446808511
from sklearn.metrics import f1_score##Calculate F1 scoreactual = []
result = []for y in y_test_merged:    if y[0] == 0:
        actual.append(0)    else:
        actual.append(1)for y in predicted_st:    if y[1] > y[0]:
        result.append(0)    else:
        result.append(1)
score = f1_score(actual,result,average='weighted',pos_label=1)
print(score)
0.498192044998
#Same percentage calculations but with a thresholdTHRESH = 0.1totalCorrect = 0total = 0noDecision = 0buyTotal = 0sellTotal = 0correct = 0sellCorrect = 0buyCorrect = 0for i in range(len(predicted_st)):
    realAnswer = y_test_merged[i][0]    if predicted_st[i][1] - THRESH > .5:
        predicted = 0 #Buy
    elif predicted_st[i][0] - THRESH > .5:
        predicted = 1 #Sell
    else:
        predicted = 2 #Pass, do not count towards percentages because you make no decision if .6>x>.4(省略)
(347, 750, 0.46266666666666667)
(190, 351, 0.5413105413105413)
(157, 399, 0.39348370927318294)
If you just predicted all Up 0.468
If you just predicted all Down 0.532
thresholds = []
totalAcc = []
positiveAcc = []
negativeAcc = []##Graph this graph of the threshold vs accuracyfor i in range(10):
    thresh = i/100.0
    totalCorrect = 0
    total = 0
    noDecision = 0
    buyTotal = 0
    sellTotal = 0
    correct = 0
    sellCorrect = 0
    buyCorrect = 0
    for i in range(len(predicted_st)):
        realAnswer = y_test_merged[i][0]        if predicted_st[i][1] - thresh > .5:
            predicted = 0 #Buy
        elif predicted_st[i][0] - thresh > .5:
            predicted = 1 #Sell(省略)    

通过测试表明,每日价格预测,外汇有更好的表现,比传统股票。因为他有更少的噪音。

原文发布于微信公众号 - 量化投资与机器学习(ZXL_LHTZ_JQXX)

原文发表时间:2017-03-23

本文参与腾讯云自媒体分享计划,欢迎正在阅读的你也加入,一起分享。

发表于

我来说两句

0 条评论
登录 后参与评论

相关文章

来自专栏AI科技评论

开发 | 如何利用微信监管你的TF训练

AI科技评论按:本文作者Coldwings,AI科技评论获其授权发布。 之前回答问题【在机器学习模型的训练期间,大概几十分钟到几小时不等,大家都会在等实验的时候...

33380
来自专栏草根专栏

Python数据分析(二): Pandas技巧 (2)

Pandas的第一部分: http://www.cnblogs.com/cgzl/p/7681974.html github地址: https://github...

32060
来自专栏生信宝典

R包ggseqlogo 绘制seq logo图

在生物信息分析中,经常会做序列分析图(sequence logo),这里的序列指的是核苷酸(DNA/RNA链中)或氨基酸(在蛋白质序列中)。sequence l...

31530
来自专栏腾讯AlloyTeam的专栏

教你用 webgl 快速创建一个小世界

Webgl的魅力在于可以创造一个自己的3D世界,但相比较canvas2D来说,除了物体的移动旋转变换完全依赖矩阵增加了复杂度,就连生成一个物体都变得很复杂……这...

1.8K00
来自专栏生信技能树

比对NR库看看物种分布【直播】我的基因组88

前面我提前了我的基因组测序数据里面的未成功比对到人类基因组上面的那些fastq序列,也用了软件把它们组装成fasta序列,这些序列的功能是未知的,可以通过比对到...

63680
来自专栏点滴积累

geotrellis使用(十五)使用Bokeh进行栅格数据可视化统计

Geotrellis系列文章链接地址http://www.cnblogs.com/shoufengwei/p/5619419.html 目录 前言 实现方案 ...

36670
来自专栏贾志刚-OpenCV学堂

手撕OpenCV源码之filter2D(一)

在上篇的GaussianBlur中提到,gaussianBlur使用的是filter2D的实现,因此上篇仅仅描述了高斯滤波器的生成细节,并没有针对滤波的计算细节...

33910
来自专栏琦小虾的Binary

OpenCV像素点邻域遍历效率比较,以及访问像素点的几种方法

OpenCV像素点邻域遍历效率比较,以及访问像素点的几种方法 前言: 以前笔者在项目中经常使用到OpenCV的算法,而大部分OpenCV的算法都需要进行遍历操作...

596100
来自专栏C#

开源免费的.NET图像即时处理的组件ImageProcessor

   承接以前的组件系列,这个组件系列旨在介绍.NET相关的组件,让大家可以在项目中有一个更好的选择组件的介绍绝对不是一篇文章可以叙述完的,因为一个组件是经过开...

33180
来自专栏Python小屋

Python切分图像小案例(1、3、2、4象限子图互换)

首先解释上一篇文章详解Python科学计算扩展库numpy中的矩阵运算(1)最后的习题,该问题答案是10 ** 8 = 100000000,原因在于Python...

43670

扫码关注云+社区

领取腾讯云代金券