深度有趣 | 10 股票價(jià)格預(yù)測(cè)

簡介股票價(jià)格預(yù)測(cè)是一件非?；Ｈ说氖虑?，但如果只基于歷史數(shù)據(jù)進(jìn)行預(yù)測(cè)，顯然完全不靠譜
股票價(jià)格是典型的時(shí)間序列數(shù)據(jù)（簡稱時(shí)序數(shù)據(jù)），會(huì)受到經(jīng)濟(jì)環(huán)境、政府政策、人為操作多種復(fù)雜因素的影響
不像氣象數(shù)據(jù)那樣具備明顯的時(shí)間和季節(jié)性模式，例如一天之內(nèi)和一年之內(nèi)的氣溫變化等
盡管如此，以股票價(jià)格為例，介紹如何對(duì)時(shí)序數(shù)據(jù)進(jìn)行預(yù)測(cè)，仍然值得一做
以下使用TensorFlow和Keras，對(duì)S&P 500股價(jià)數(shù)據(jù)進(jìn)行分析和預(yù)測(cè)
數(shù)據(jù)S&P 500股價(jià)數(shù)據(jù)爬取自Google Finance API，已經(jīng)進(jìn)行過缺失值處理
加載庫，pandas主要用于數(shù)據(jù)清洗和整理
# -*- coding: utf-8 -*-import pandas as pdimport numpy as npimport tensorflow as tfimport matplotlib.pyplot as plt%matplotlib inlinefrom sklearn.preprocessing import MinMaxScalerimport time復(fù)制代碼用pandas讀取csv文件為DataFrame，并用describe()查看特征的數(shù)值分布
data = pd.read_csv('data_stocks.csv')data.describe()復(fù)制代碼還可以用info()查看特征的概要
data.info()復(fù)制代碼數(shù)據(jù)共502列，41266行，502列分別為：
DATE：該行數(shù)據(jù)的時(shí)間戳SP500：可以理解為大盤指數(shù)其他：可以理解為500支個(gè)股的股價(jià)
查看數(shù)據(jù)的前五行
data.head()復(fù)制代碼查看時(shí)間跨度
print(time.strftime('%Y-%m-%d', time.localtime(data['DATE'].max())),       time.strftime('%Y-%m-%d', time.localtime(data['DATE'].min())))復(fù)制代碼繪制大盤趨勢(shì)折線圖
plt.plot(data['SP500'])復(fù)制代碼去掉DATE一列，訓(xùn)練集測(cè)試集分割
data.drop('DATE', axis=1, inplace=True)data_train = data.iloc[:int(data.shape[0] * 0.8), :]data_test = data.iloc[int(data.shape[0] * 0.8):, :]print(data_train.shape, data_test.shape)復(fù)制代碼數(shù)據(jù)歸一化，只能使用data_train進(jìn)行fit()
scaler = MinMaxScaler(feature_range=(-1, 1))scaler.fit(data_train)data_train = scaler.transform(data_train)data_test = scaler.transform(data_test)復(fù)制代碼同步預(yù)測(cè)同步預(yù)測(cè)是指，使用當(dāng)前時(shí)刻的500支個(gè)股股價(jià)，預(yù)測(cè)當(dāng)前時(shí)刻的大盤指數(shù)，即一個(gè)回歸問題，輸入共500維特征，輸出一維，即[None, 500] => [None, 1]
使用TensorFlow實(shí)現(xiàn)同步預(yù)測(cè)，主要用到多層感知機(jī)（Multi-Layer Perceptron，MLP），損失函數(shù)用均方誤差（Mean Square Error，MSE）
X_train = data_train[:, 1:]y_train = data_train[:, 0]X_test = data_test[:, 1:]y_test = data_test[:, 0]input_dim = X_train.shape[1]hidden_1 = 1024hidden_2 = 512hidden_3 = 256hidden_4 = 128output_dim = 1batch_size = 256epochs = 10tf.reset_default_graph()X = tf.placeholder(shape=[None, input_dim], dtype=tf.float32)Y = tf.placeholder(shape=[None], dtype=tf.float32)W1 = tf.get_variable('W1', [input_dim, hidden_1], initializer=tf.contrib.layers.xavier_initializer(seed=1))b1 = tf.get_variable('b1', [hidden_1], initializer=tf.zeros_initializer())W2 = tf.get_variable('W2', [hidden_1, hidden_2], initializer=tf.contrib.layers.xavier_initializer(seed=1))b2 = tf.get_variable('b2', [hidden_2], initializer=tf.zeros_initializer())W3 = tf.get_variable('W3', [hidden_2, hidden_3], initializer=tf.contrib.layers.xavier_initializer(seed=1))b3 = tf.get_variable('b3', [hidden_3], initializer=tf.zeros_initializer())W4 = tf.get_variable('W4', [hidden_3, hidden_4], initializer=tf.contrib.layers.xavier_initializer(seed=1))b4 = tf.get_variable('b4', [hidden_4], initializer=tf.zeros_initializer())W5 = tf.get_variable('W5', [hidden_4, output_dim], initializer=tf.contrib.layers.xavier_initializer(seed=1))b5 = tf.get_variable('b5', [output_dim], initializer=tf.zeros_initializer())h1 = tf.nn.relu(tf.add(tf.matmul(X, W1), b1))h2 = tf.nn.relu(tf.add(tf.matmul(h1, W2), b2))h3 = tf.nn.relu(tf.add(tf.matmul(h2, W3), b3))h4 = tf.nn.relu(tf.add(tf.matmul(h3, W4), b4))out = tf.transpose(tf.add(tf.matmul(h4, W5), b5))cost = tf.reduce_mean(tf.squared_difference(out, Y))optimizer = tf.train.AdamOptimizer().minimize(cost)with tf.Session() as sess:    sess.run(tf.global_variables_initializer())    for e in range(epochs):        shuffle_indices = np.random.permutation(np.arange(y_train.shape[0]))        X_train = X_train[shuffle_indices]        y_train = y_train[shuffle_indices]        for i in range(y_train.shape[0] // batch_size):            start = i * batch_size            batch_x = X_train[start : start + batch_size]            batch_y = y_train[start : start + batch_size]            sess.run(optimizer, feed_dict={X: batch_x, Y: batch_y})            if i % 50 == 0:                print('MSE Train:', sess.run(cost, feed_dict={X: X_train, Y: y_train}))                print('MSE Test:', sess.run(cost, feed_dict={X: X_test, Y: y_test}))                y_pred = sess.run(out, feed_dict={X: X_test})                y_pred = np.squeeze(y_pred)                plt.plot(y_test, label='test')                plt.plot(y_pred, label='pred')                plt.title('Epoch ' + str(e) + ', Batch ' + str(i))                plt.legend()                plt.show()復(fù)制代碼最后測(cè)試集的loss在0.005左右，預(yù)測(cè)結(jié)果如下    

使用Keras實(shí)現(xiàn)同步預(yù)測(cè)，代碼量會(huì)少很多，但具體實(shí)現(xiàn)細(xì)節(jié)不及TensorFlow靈活
from keras.layers import Input, Densefrom keras.models import ModelX_train = data_train[:, 1:]y_train = data_train[:, 0]X_test = data_test[:, 1:]y_test = data_test[:, 0]input_dim = X_train.shape[1]hidden_1 = 1024hidden_2 = 512hidden_3 = 256hidden_4 = 128output_dim = 1batch_size = 256epochs = 10X = Input(shape=[input_dim,])h = Dense(hidden_1, activation='relu')(X)h = Dense(hidden_2, activation='relu')(h)h = Dense(hidden_3, activation='relu')(h)h = Dense(hidden_4, activation='relu')(h)Y = Dense(output_dim, activation='sigmoid')(h)model = Model(X, Y)model.compile(loss='mean_squared_error', optimizer='adam')model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size, shuffle=False)y_pred = model.predict(X_test)print('MSE Train:', model.evaluate(X_train, y_train, batch_size=batch_size))print('MSE Test:', model.evaluate(X_test, y_test, batch_size=batch_size))plt.plot(y_test, label='test')plt.plot(y_pred, label='pred')plt.legend()plt.show()復(fù)制代碼最后測(cè)試集的loss在0.007左右，預(yù)測(cè)結(jié)果如下

異步預(yù)測(cè)異步預(yù)測(cè)是指，使用歷史若干個(gè)時(shí)刻的大盤指數(shù)，預(yù)測(cè)當(dāng)前時(shí)刻的大盤指數(shù)，這樣才更加符合預(yù)測(cè)的定義
例如，使用前五個(gè)大盤指數(shù)，預(yù)測(cè)當(dāng)前的大盤指數(shù)，每組輸入包括5個(gè)step，每個(gè)step對(duì)應(yīng)一個(gè)歷史時(shí)刻的大盤指數(shù)，輸出一維，即[None, 5, 1] => [None, 1]
使用Keras實(shí)現(xiàn)異步預(yù)測(cè)，主要用到循環(huán)神經(jīng)網(wǎng)絡(luò)即RNN（Recurrent Neural Network）中的LSTM（Long Short-Term Memory）
from keras.layers import Input, Dense, LSTMfrom keras.models import Modeloutput_dim = 1batch_size = 256epochs = 10seq_len = 5hidden_size = 128X_train = np.array([data_train[i : i + seq_len, 0] for i in range(data_train.shape[0] - seq_len)])[:, :, np.newaxis]y_train = np.array([data_train[i + seq_len, 0] for i in range(data_train.shape[0] - seq_len)])X_test = np.array([data_test[i : i + seq_len, 0] for i in range(data_test.shape[0] - seq_len)])[:, :, np.newaxis]y_test = np.array([data_test[i + seq_len, 0] for i in range(data_test.shape[0] - seq_len)])print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)X = Input(shape=[X_train.shape[1], X_train.shape[2],])h = LSTM(hidden_size, activation='relu')(X)Y = Dense(output_dim, activation='sigmoid')(h)model = Model(X, Y)model.compile(loss='mean_squared_error', optimizer='adam')model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size, shuffle=False)y_pred = model.predict(X_test)print('MSE Train:', model.evaluate(X_train, y_train, batch_size=batch_size))print('MSE Test:', model.evaluate(X_test, y_test, batch_size=batch_size))plt.plot(y_test, label='test')plt.plot(y_pred, label='pred')plt.legend()plt.show()復(fù)制代碼最后測(cè)試集的loss在0.0015左右，預(yù)測(cè)結(jié)果如下，一層LSTM的效果已經(jīng)好非常多了

 當(dāng)然，還有一種可能的嘗試，使用歷史若干個(gè)時(shí)刻的500支個(gè)股股價(jià)以及大盤指數(shù)，預(yù)測(cè)當(dāng)前時(shí)刻的大盤指數(shù)，即[None, 5, 501] => [None, 1]
from keras.layers import Input, Dense, LSTMfrom keras.models import Modeloutput_dim = 1batch_size = 256epochs = 10seq_len = 5hidden_size = 128X_train = np.array([data_train[i : i + seq_len, :] for i in range(data_train.shape[0] - seq_len)])y_train = np.array([data_train[i + seq_len, 0] for i in range(data_train.shape[0] - seq_len)])X_test = np.array([data_test[i : i + seq_len, :] for i in range(data_test.shape[0] - seq_len)])y_test = np.array([data_test[i + seq_len, 0] for i in range(data_test.shape[0] - seq_len)])print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)X = Input(shape=[X_train.shape[1], X_train.shape[2],])h = LSTM(hidden_size, activation='relu')(X)Y = Dense(output_dim, activation='sigmoid')(h)model = Model(X, Y)model.compile(loss='mean_squared_error', optimizer='adam')model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size, shuffle=False)y_pred = model.predict(X_test)print('MSE Train:', model.evaluate(X_train, y_train, batch_size=batch_size))print('MSE Test:', model.evaluate(X_test, y_test, batch_size=batch_size))plt.plot(y_test, label='test')plt.plot(y_pred, label='pred')plt.legend()plt.show()復(fù)制代碼最后的loss在0.004左右，結(jié)果反而變差了
500支個(gè)股加上大盤指數(shù)的預(yù)測(cè)效果，還不如僅使用大盤指數(shù)
說明特征并不是越多越好，有時(shí)候反而會(huì)引入不必要的噪音
由于并未涉及到復(fù)雜的CNN或RNN，所以在CPU上運(yùn)行的速度還可以

本文版權(quán)歸黑馬程序員JavaEE學(xué)院所有，歡迎轉(zhuǎn)載，轉(zhuǎn)載請(qǐng)注明作者出處。謝謝！

作者：黑馬程序員python+人工智能培訓(xùn)學(xué)院

首發(fā)：http://python.itheima.com/?v2