【tensorflow】利用神经网络绘制股票价格拟合曲线

神经网络有三层，输入层A，隐藏层B，输出层C，满足：

A(10x1)*W1(1x10)+b1(1x10)=B(10x10)

B(10x10)*W2(10x1)+b2(10x1)=C(10x1)

我们需要做的，便是通过多次训练(尝试不同 w、b 的值)，找到合适的 w1w2、b1b2，使预测结果更接近真实结果。

代码：

import tensorflow.compat.v1 as tf
import numpy as np
import matplotlib.pyplot as plt

# 兼容 tensorflow 1.0
tf.disable_eager_execution()

# 日期，0-9
date = np.linspace(0, 9, 10)
# 收盘价格
endPrice = np.array([2511.90, 2538.26, 2510.68, 2591.66, 2732.98, 2701.69, 2701.29, 2678.67, 2726.50, 2681.50])
# 开盘价格
beginPrice = np.array([2438.71, 2739.17, 2715.07, 2823.58, 2864.90, 2919.08, 2500.88, 2534.95, 2512.52, 2594.04])

# 绘制涨幅趋势
# 价格上涨，红色
# 价格下跌，绿色
plt.figure()
for i in range(10):
    temData = np.zeros([2])
    temData[0] = i
    temData[1] = i
    temPrice = np.zeros([2])
    temPrice[0] = beginPrice[i]
    temPrice[1] = endPrice[i]
    if beginPrice[i] > endPrice[i]:
        plt.plot(temData, temPrice, "green", lw="2")
    else:
        plt.plot(temData, temPrice, "red", lw="2")

# 神经网络有三层，输入层A，隐藏层B，输出层C
# A(10x1)*W1(1x10)+b1(1x10)=B(10x10)
# B(10x10)*W2(10x1)+b2(10x1)=C(10x1)

# 输入层
# 归一化处理，减小计算量
dateNormal = np.zeros([10, 1])
priceNormal = np.zeros(([10, 1]))
for i in range(10):
    dateNormal[i] = i / 9.0
    priceNormal[i] = endPrice[i]/3000.0
    
# 声明输入输出参数，xy
x = tf.placeholder(tf.float32, [None, 1])
y = tf.placeholder(tf.float32, [None, 1])

# 隐藏层
w1 = tf.Variable(tf.random_uniform([1, 10], 0, 1))
b1 = tf.Variable(tf.zeros([1, 10]))
wb1 = tf.matmul(x, w1) + b1
# 激励函数
layer1 = tf.nn.relu(wb1)

# 输出层
w2 = tf.Variable(tf.random_uniform([10, 1], 0, 1))
b2 = tf.Variable(tf.zeros([10, 1]))
wb2 = tf.matmul(layer1, w2) + b2
# 激励函数
layer2 = tf.nn.relu(wb2)

# 预测值与实际值的标准差
loss = tf.reduce_mean(tf.square(y - layer2))

# 使用梯度下降法，减小差异
tran_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)

# 变量初始化
init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    
    # 训练10000次
    for i in range(10000):
        sess.run(tran_step, feed_dict={x: dateNormal, y: priceNormal})
    
    # 查看拟合结果
    pre = sess.run(layer2, feed_dict={x: dateNormal})
    prePrice = np.zeros([10, 1])
    # 反归一化
    for i in range(10):
        prePrice[i, 0] = (pre * 3000)[i, 0]
        
    # 绘制折线
    plt.plot(date, prePrice, "blue", lw=2)
    
plt.show()

运行效果：