**Notes before writing:**

A few days ago, I wore a thin coat for a few days, but I was defeated by the weather in Guangzhou (too hot). So it’s still short sleeve shorts. It’s not cold in Guangzhou.

#### Environmental preparation

Tensorflow 2.3.1 was used in this experiment.

```
import numpy as np
import matplotlib.pyplot as plt
from pandas import read_csv
import math
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_squared_error
%matplotlib inline
```

#### data description

Using the historical stock price data of Yahoo! Finance ^ GSPC in the past five years, from November 2015 to November 2020, a total of 1256. This data contains daily stock price information, such as date, open, high, low, close, adj close, volume.

*Tips: stock knowledge*

*Date: Date**Open: opening price (the starting price of a stock on a certain day)**High: the highest price**Low: the lowest price**Close: closing price (the final price of the stock on a certain day)**Adj close: weighted closing price**Volume: total transaction volume*

For simplicity, only the closing price is used for forecasting. The chart below shows the closing price of the past five years.

code:

```
#Loading dataset with pandas
dataframe = read_csv('data/stock_data.csv', usecols=[4], engine='python', skipfooter=3)
data = dataframe.values
#Change integer to float
data = data.astype('float32')
plt.plot(data)
plt.show()
```

#### target

Forecast the future stock closing price, this forecast is the last 56 data.

#### Build training set and test set

The closing price of the past five years is a length of*N*The time series of P_{0}, p_{1},…,p_{N-1}For the price of each day. Before use*i*Data forecast No*i*+1 data construction training set and test set, 0<*i* < *N*, i.e

X_{0}= (p_{0}, p_{1},…, p_{i-1})X_{1}= (p_{i}, p_{i+1},…, p_{2i-1})

…X_{t}= (p_{ti}, p_{ti+1},…, p_{(t+1)i-1})

To predict

X_{t+1}= (p_{(t+1)i}, p_{(t+1)i+1},…, p_{(t+2)i-1})

Choose here*i*= 6。 In LSTM, time_ If steps = 6, the training set can be expressed as

Input

_{1}= [p_{0}, p_{1}, p_{2}, p_{3}, p_{4}, p_{5}], Label_{1}= [p_{6}]

Input_{2}= [p_{1}, p_{2}, p_{3}, p_{4}, p_{5}, p_{6}], Label_{1}= [p_{7}]

Input_{3}= [p_{2}, p_{3}, p_{4}, p_{5}, p_{6}, p_{7}], Label_{1}= [p_{8}]

code:

```
#Construct matrix from original data set
def create_dataset(data, time_steps):
dataX, dataY = [], []
for i in range(len(data) - time_steps):
a = data[i:(i + time_steps), 0]
dataX.append(a)
dataY.append(data[i + time_steps, 0])
return np.array(dataX), np.array(dataY)
```

Set 95.55% as training set and the rest as test set

```
#Normalization
scaler = MinMaxScaler(feature_range=(0, 1))
data = scaler.fit_transform(data)
#Cut into training set and test set
train_size = int(len(data) * 0.9555)
test_size = len(data) - train_size
train, test = data[0:train_size,:], data[train_size:len(data),:]
time_steps = 6
trainX, trainY = create_dataset(train, time_steps)
testX, testY = create_dataset(test, time_steps)
#The format of reshape input model data is: [samples, time steps, features]
trainX = np.reshape(trainX, (trainX.shape[0], trainX.shape[1], 1))
testX = np.reshape(testX, (testX.shape[0], testX.shape[1], 1))
```

#### Establish and train LSTM model

The number of neurons in the hidden layer is 128, the output layer is 1 predictive value, and the number of iterations is 100.

*Tips: Calculation of LSTM parameters*

(hidden size × (hidden size + x_dim) + hidden size) × 4

X_ Dim is the characteristic dimension of input data, here is 1.

code:

```
model = Sequential()
model.add(LSTM(128, input_shape=(time_steps, 1)))
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])
model.summary()
history = model.fit(trainX, trainY, epochs=100, batch_size=64, verbose=1)
score = model.evaluate(testX, testY, batch_size=64, verbose=1)
```

The result of loss function of visual training set is shown in the figure below. It can be seen that the value of loss converges gradually.

code:

```
def visualize_loss(history, title):
loss = history.history["loss"]
epochs = range(len(loss))
plt.figure()
plt.plot(epochs, loss, "b", label="Training loss")
plt.title(title)
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.legend()
plt.show()
visualize_loss(history, "Training Loss")
```

#### Forecast results

code:

```
#Prediction training set and test set
trainPredict = model.predict(trainX)
testPredict = model.predict(testX)
#The prediction results are inverse normalized
trainPredict = scaler.inverse_transform(trainPredict)
trainY = scaler.inverse_transform([trainY])
testPredict = scaler.inverse_transform(testPredict)
testY = scaler.inverse_transform([testY])
#Calculating RMSE of training set and test set
trainScore = math.sqrt(mean_squared_error(trainY[0], trainPredict[:,0]))
print('Train Score: %.2f RMSE' % (trainScore))
testScore = math.sqrt(mean_squared_error(testY[0], testPredict[:,0]))
print('Test Score: %.2f RMSE' % (testScore))
#Draw the prediction result graph
trainPredictPlot = np.empty_like(data)
trainPredictPlot[:, :] = np.nan
trainPredictPlot[time_steps:len(trainPredict) + time_steps, :] = trainPredict
testPredictPlot = np.empty_like(data)
testPredictPlot[:, :] = np.nan
testPredictPlot[len(trainPredict) + (time_steps * 2)-1:len(data) - 1, :] = testPredict
plt.plot(scaler.inverse_transform(data))
plt.plot(trainPredictPlot)
plt.plot(testPredictPlot)
plt.show()
```

In the above figure, the blue line is the original data, the orange line and the green line are the prediction results of the training set and the test set, respectively.

#### reference resources

https://www.jianshu.com/p/38d…

https://keras.io/examples/tim…