如何使用 LSTM 模型进行多步预测？

Question

我使用 LSTM 开发了一个时间序列模型。我不能用它来预测未来几天的股价。我想用它来预测明年的股价并绘制它。如何用它来预测未来（明年）的股价？

df=pd.read_csv('foolad.csv')
df=df.set_index(pd.DatetimeIndex(df['Date'].values))

data=df.filter(['Close'])
dataset=data.values

training_data_len=math.ceil(len(dataset)*0.8)
scaler=MinMaxScaler(feature_range=(0,1))
scaled_data=scaler.fit_transform(dataset)
scaled_data

training_data=scaled_data[0:training_data_len , :]

xtrain=[]
ytrain=[]
n = 60

for i in range(n,len(training_data)):
    xtrain.append(training_data[i-n:i , 0])
    ytrain.append(training_data[i,0])

xtrain , ytrain = np.array(xtrain) , np.array(ytrain)
xtrain=np.reshape(xtrain , (xtrain.shape[0],xtrain.shape[1],1))
xtrain.shape

model=Sequential()
model.add(LSTM(50,return_sequences=True,input_shape=(xtrain.shape[1],1)))
model.add(LSTM(50,return_sequences=False))
model.add(Dense(25))
model.add(Dense(1))

model.compile(loss='mean_squared_error',optimizer='adam')

model.fit(xtrain,ytrain,epochs=1,batch_size=1)

test_data=scaled_data[training_data_len - n : , :]
xtest=[]
ytest=dataset[training_data_len : , :]
for i in range(n , len(test_data)):
    xtest.append(test_data[i-n : i , 0])

xtest=np.array(xtest)
xtest=np.reshape(xtest , (xtest.shape[0],xtest.shape[1],1))

prediction=model.predict(xtest)
prediction=scaler.inverse_transform(prediction)

#for future 360 days what can I do?....

Answer 1

一种方法是将预测作为输入反馈给模型：在每个步骤中，您通过删除最旧的值并添加最新的预测作为最新值来更新输入序列。如下图所示，其中 n 是输入序列的长度，T 是时间序列的长度。

下面的代码展示了如何为 LSTM 模型实现这种方法并绘制结果。

import numpy as np
import pandas as pd
import yfinance as yf
import tensorflow as tf
from tensorflow.keras.layers import Dense, LSTM
from tensorflow.keras.models import Sequential
from sklearn.preprocessing import MinMaxScaler

# download the data
df = yf.download(tickers=['^IXIC'], period='5y')
y = df['Close'].fillna(method='ffill').values.reshape(- 1, 1)

# scale the data
scaler = MinMaxScaler(feature_range=(0, 1))
scaler = scaler.fit(y)
y = scaler.transform(y)

# generate the training sequences
n_forecast = 1 
n_lookback = 60

X = []
Y = []

for i in range(n_lookback, len(y) - n_forecast + 1):
    X.append(y[i - n_lookback: i])
    Y.append(y[i: i + n_forecast])

X = np.array(X)
Y = np.array(Y)

# train the model
tf.random.set_seed(0)

model = Sequential()
model.add(LSTM(50, return_sequences=True, input_shape=(X.shape[1], X.shape[2])))
model.add(LSTM(50, return_sequences=False))
model.add(Dense(25))
model.add(Dense(1))

model.compile(loss='mse', optimizer='adam')
model.fit(X, Y, epochs=100, batch_size=128, validation_split=0.2, verbose=0)

# generate the multi-step forecasts
n_future = 365
y_future = []

x_pred = X[-1:, :, :]  # last observed input sequence
y_pred = Y[-1]         # last observed target value

for i in range(n_future):

    # feed the last forecast back to the model as an input
    x_pred = np.append(x_pred[:, 1:, :], y_pred.reshape(1, 1, 1), axis=1)

    # generate the next forecast
    y_pred = model.predict(x_pred)

    # save the forecast
    y_future.append(y_pred.flatten()[0])

# transform the forecasts back to the original scale
y_future = np.array(y_future).reshape(-1, 1)
y_future = scaler.inverse_transform(y_future)

# organize the results in a data frame
df_past = df[['Close']].reset_index()
df_past.rename(columns={'index': 'Date'}, inplace=True)
df_past['Date'] = pd.to_datetime(df_past['Date'])
df_past['Forecast'] = np.nan

df_future = pd.DataFrame(columns=['Date', 'Close', 'Forecast'])
df_future['Date'] = pd.date_range(start=df_past['Date'].iloc[-1] + pd.Timedelta(days=1), periods=n_future)
df_future['Forecast'] = y_future.flatten()
df_future['Close'] = np.nan

results = df_past.append(df_future).set_index('Date')

# plot the results
results.plot(title='NASDAQ')