Stock price forecast with tensorflow

http://qiita.com/northriver/items/4f4690053e1770311335

I'm used to tensorlfow, so I'll try to predict the stock price by deep learning Create a record date and use the stock price of the past 90 days (on business days) and the stock price 30 days later to put out a stock that is likely to rise.

The correct answer data after 30 days was selected as x% higher, lower, or in the meantime with respect to the standard. I decided to solve it as a classification problem.

First, make the necessary functions with two hidden layers

import tensorflow as tf
import numpy as np
sess = tf.InteractiveSession()


def inference(x_placeholder,SCORE_SIZE,NUM_HIDDEN1,NUM_OUTPUT,keep_prob):
	with tf.name_scope('hidden1') as scope:
		hidden1_weight = tf.Variable(tf.truncated_normal([SCORE_SIZE, NUM_HIDDEN1], stddev=0.1), name="hidden1_weight")
		hidden1_bias = tf.Variable(tf.constant(0.1, shape=[NUM_HIDDEN1]), name="hidden1_bias")
		mean1, variance1 = tf.nn.moments(x_placeholder,[0])
		bn1 = tf.nn.batch_normalization(x_placeholder, mean1, variance1, None, None, 1e-5)
		hidden1_output = tf.nn.relu(tf.matmul(bn1, hidden1_weight) + hidden1_bias)

	with tf.name_scope('hidden2'):
		hidden2_weight = tf.Variable(tf.truncated_normal([NUM_HIDDEN1, NUM_HIDDEN2], stddev=0.01), name='hidden2_weight')
		hidden2_bias = tf.Variable(tf.constant(0.1, shape=[NUM_HIDDEN2]), name="hidden2_bias")
		mean2, variance2 = tf.nn.moments(hidden1_output,[0])
		bn2 = tf.nn.batch_normalization(hidden1_output, mean2, variance2, None, None, 1e-5)
		hidden2_output = tf.nn.relu(tf.matmul(bn2, hidden2_weight) + hidden2_bias)

	dropout = tf.nn.dropout(hidden1_output, keep_prob)

	with tf.name_scope('output') as scope:
		output_weight = tf.Variable(tf.truncated_normal([NUM_HIDDEN1, NUM_OUTPUT], stddev=0.1), name="output_weight")
		output_bias = tf.Variable(tf.constant(0.1, shape=[NUM_OUTPUT]), name="output_bias")
		mean_out, variance_out = tf.nn.moments(dropout,[0])
		bn_out = tf.nn.batch_normalization(dropout, mean_out, variance_out, None, None, 1e-5)
		output = tf.nn.softmax(tf.matmul(bn_out, output_weight) + output_bias)
	return output

def loss(output, y_placeholder, loss_label_placeholder):
	with tf.name_scope('loss') as scope:
		loss = tf.reduce_mean(-tf.reduce_sum(y_placeholder * tf.log(tf.clip_by_value(output,1e-10,1.0)), reduction_indices=[1]))
	return loss

def training(loss,rate):
	with tf.name_scope('training') as scope:
		train_step = tf.train.GradientDescentOptimizer(rate).minimize(loss)
	return train_step

Here, I was careful that it is better to do batch normalization, so I tried to dig into all layers

The number of hidden layers and the learning rate are appropriate, and will be adjusted in the future. Run!

rate=0.08
SCORE_SIZE=90
NUM_OUTPUT=3
NUM_HIDDEN1=90
NUM_HIDDEN2=45


y_placeholder = tf.placeholder("float", [None, NUM_OUTPUT], name="y_placeholder")
x_placeholder = tf.placeholder("float", [None, SCORE_SIZE], name="x_placeholder")
loss_label_placeholder = tf.placeholder("string", name="loss_label_placeholder")
keep_prob = tf.placeholder("float")


output = inference(x_placeholder,SCORE_SIZE,NUM_HIDDEN1,NUM_OUTPUT,keep_prob)
loss1 = loss(output, y_placeholder, loss_label_placeholder)
training_op = training(loss1,rate)

correct_prediction = tf.equal(tf.argmax(output,1), tf.argmax(y_placeholder,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

best_loss = float("inf")

sess.run(tf.initialize_all_variables())
for step in range(1001):
	x_train,y_train,c_train=train_data_make(SCORE_SIZE,200)
	feed_dict_train={y_placeholder: y_train,x_placeholder: x_train,keep_prob: 0.8}
	feed_dict_test={y_placeholder: y_test,x_placeholder: x_test, keep_prob: 1.0}

	if step%100 == 0:print "size %d,step %d,rate %g loss %g,test accuracy %g"%(num_long,step,rate,loss1.eval(feed_dict=feed_dict_train),accuracy.eval(feed_dict=feed_dict_test))
	sess.run(training_op, feed_dict=feed_dict_train)	

#test_result
print "final_result : size %d,step %d, test accuracy %g"%(num_long,step,accuracy.eval(feed_dict=feed_dict_test))

As a result, 9 are allocated ... Well, most of the stocks have the same stock price in one month, so it may be due to no change. Rather, 90% of the test data was classified without change ...

size 90,step 0,rate 0.08 loss 1.31044,test accuracy 0.257594
size 90,step 100,rate 0.08 loss 0.369047,test accuracy 0.91373
size 90,90,step 200,rate 0.08 loss 0.249306,test accuracy 0.90887
・ ・ ・

Let's take a look at what kind of numbers are given for each stock. After all, there were many expectations that it wouldn't change, but ... that? If you arrange them in descending order of likelihood of increase, some stocks are expected to do so.

Is it also a unique movement before the stock price goes up?

I would like to see the situation for a while and observe if it really goes up.

According to the following, it seems that the batch size should be smaller http://postd.cc/26-things-i-learned-in-the-deep-learning-summer-school/ (Invest in stocks at your own risk. Articles are not responsible at all)