In this paper , We will use neural networks , especially LSTM Model to predict the behavior of time series data . The problem to be solved is the classical stock market forecast . All the data and codes used can be trusted to me .
Although this is an old problem , But it has not been solved until today . The fact is simple ： Stock prices are determined by several factors , Historical stock prices are only a small part of it . therefore , Forecasting price behavior is a very difficult problem .
abstract
first , I will introduce some data visualization
Data set for . then , I will briefly discuss the difficulty and limitations of using moving average algorithm to predict stock market behavior . next , Recursive neural networks and LSTM Concept of , And with LSTM For example, for a single company
The stock price is predicted . last , I'll show you how to predict at the same time 4 Company price LSTM, And compare the results , See as we use more companies at the same time , Does the forecast improve .
Data visualization
The dataset is from Yahoo
Finance with CSV Downloaded in format . It has 4 The stock prices of companies are 01/08/2010 to 01/07/2019 period . We call them companies A,B,C and D.
The basic step is to use Pandas open CSV file . First look at the data ：
df_A = pd.read_csv(‘data/company_A.csv’)df_A[‘Date’] =
pd.to_datetime(df_A[‘Date’])df_A.tail() Plt.figure(figsize =
(15,10))plt.plot(df_A['Date'], df_A['Close'], label='Company
A')plt.plot(df_B['Date'], df_B['Close'], label='Company
B')plt.plot(df_C['Date'], df_C['Close'], label='Company
C')plt.plot(df_D['Date'], df_D['Close'], label='Company
D')plt.legend(loc='best')plt.show()
4 The closing price of all the company's shares
Moving average
A classical algorithm used in this problem is moving average (MA). It includes calculation m Average of past observation days , And use this result as the prediction of the next day . To prove this , Here is an example of a moving average , use m As the closing price of the company 10 and 20 day .
df['MA_window_10'] = df['Close'].rolling(10).mean().shift() #shift so the day
we want to predict won't be useddf['MA_window_20'] =
df['Close'].rolling(20).mean().shift()
Use moving average pairs A Make a one-step forecast of the company's closing price
When we try to use the moving average to predict the future 10 Day closing price , give the result as follows ：
Use moving average to company A of 10 Day closing price forecast
Each red line represents based on the past 10 Day forecast . therefore , The red line is discontinuous .
Use exponential moving average (EMA), We have achieved some small improvements ：
Use the index moving average to the company A Make a one-step prediction of the closing price of .
contrast MA and EMA：
use MA and EMA yes A Comparison of one-step prediction of company closing price
This method is simple . What we really want is advance n Forecast the future trend of the stock ,MA and EMA Can't complete the task .
Recurrent neural network (RNN)
To understand LSTM network , We first need to understand recurrent neural networks . When past results have an impact on current results , This network is used to identify patterns .RNN An example used is the time series function , The data order is very important .
In this network architecture , Neurons use not only conventional inputs ( Previous layer output ) As input , It also uses its previous status as input .
RNN framework
it is to be noted that H Represents neuronal state . therefore , When in state H_1 Time , Neuron usage parameters X_1 and H_0( Its previous state ) As input . The main problem with this model is memory loss . The old state will soon be forgotten . In the sequence we need to remember just in the past ,RNN Can't remember .
LSTM network
LSTM Network originated from RNN. But it can solve memory loss by changing the structure of neurons .
LSTM Neuronal architecture
New neurons have 3 A door , Each door has a different goal . The door is ：
* Input gate
* Output gate
* Forget the door
LSTM Neurons still receive their previous state as input ：
LSTM neuron n Pass its previous state as a parameter .
LSTM Forecast individual companies
last , Let's use LSTM To predict the company A act .
But first , Consider the following parameters . We want to predict the future n day (foward_days) Enter past observations m day (look_back). therefore , If we have a past m Day input , Network output will be the future n Day forecast . We will Train and Test Split data in . The test will be conducted by k Cycles (num_period) form , Each cycle is a series of n Day forecast .
look_back = 40forward_days = 10num_periods = 20
Now? , We use Pandas open CSV file , Keep only the columns we will use , Date and closing price .A The company's initial closing price chart is ：
plt.figure(figsize = (15,10))plt.plot(df)plt.show()
A Closing price of the company
In order , We scale the input , stay Train/Validation and Test Split data in , And format it to provide a model .
Now? , We build and train models .
NUM_NEURONS_FirstLayer = 128NUM_NEURONS_SecondLayer = 64EPOCHS = 220#Build the
modelmodel =
Sequential()model.add(LSTM(NUM_NEURONS_FirstLayer,input_shape=(look_back,1),
return_sequences=True))model.add(LSTM(NUM_NEURONS_SecondLayer,input_shape=(NUM_NEURONS_FirstLayer,1)))model.add(Dense(foward_days))model.compile(loss='mean_squared_error',
optimizer='adam')history =
model.fit(X_train,y_train,epochs=EPOCHS,validation_data=(X_validate,y_validate),shuffle=True,batch_size=2,
verbose=2)
The result is ：
A The company's model applies to all data
Observe the test set carefully ：
A The company's model is applicable to the test set
Please note that , Each red line represents a based on the past 40 day (look_back) of 10 Day forecast (forward -
days). have 20 Red line , Because we chose to 20 Cycles (num_periods) Test on . This is why the red prediction line is discontinuous .
By repeating the same process for all companies , The best results on the test set are for the company C Prediction of ：
C The company's model is applicable to the test set
Although this is the best model , But the result is not very good . There are many possible reasons for this result . Some of them may be ：
* Only the historical data of the closing price is not enough to predict the stock price
* The model can also be improved
LSTM forecast 4 Company
last , We will use LSTM The model predicts all 4 The behavior of this company ,A,B,C and D, And with a single LSTM Contrast the company's results . The purpose is to analyze whether the use of data from multiple companies can improve individual forecasts for each company .
It should be pointed out that , All 4 individual csv All have the same date . such , The network will not receive future information from one company to predict the value of another company .
The initial data is ：
All 4 Closing price of the company
After data normalization and formatting of the model , Train the model ：
NUM_NEURONS_FirstLayer = 100NUM_NEURONS_SecondLayer = 50EPOCHS = 200#Build the
modelmodel =
Sequential()model.add(LSTM(NUM_NEURONS_FirstLayer,input_shape=(look_back,num_companies),
return_sequences=True))model.add(LSTM(NUM_NEURONS_SecondLayer,input_shape=(NUM_NEURONS_FirstLayer,1)))model.add(Dense(foward_days
* num_companies))model.compile(loss='mean_squared_error',
optimizer='adam')history =
model.fit(X_train,y_train,epochs=EPOCHS,validation_data=(X_validate,y_validate),shuffle=True,batch_size=1,
verbose=2)
The result is ：
4 Company LSTM Results of the model
Observe the test set carefully ：
4 Company LSTM Test set prediction of model
It's time to compare the results . Single company LSTM The results are displayed on the left ,4 Companies LSTM The results are displayed on the right . The first row shows the predictions in the test set , The second row shows the forecasts in all data sets .
A company
B company
C company
D company
conclusion
It is impossible to predict the behavior of the stock market only by historical prices .LSTM The forecast is unacceptable . Even using the historical prices of several companies , Predictions are getting worse .
Technology