lylayang

from statsmodels.tsa.api import ExponentialSmoothing
fit1 = ExponentialSmoothing(np.asarray(train['Spend']) ,seasonal_periods=7 ,trend='add', seasonal='add').fit(use_boxcox=True)
test['Holt_Winter'] = fit1.forecast(len(test))
plt.figure(figsize=(16,8))
plt.plot( train['Spend'], label='Train')
plt.plot(test['Spend'], label='Test')
plt.plot(test['Holt_Winter'], label='Holt_Winter')
plt.legend(loc='best')
plt.show()

lylayang

import statsmodels.api as sm
fit1 = sm.tsa.statespace.SARIMAX(train.Spend, order=(7, 1, 2), seasonal_order=(0, 1, 2, 7)).fit(use_boxcox=True)
test['SARIMA'] = fit1.predict(start="2019-07-23", end="2019-09-23", dynamic=True)
plt.figure(figsize=(16, 8))
plt.plot(train['Spend'], label='Train')
plt.plot(test['Spend'], label='Test')
plt.plot(test['SARIMA'], label='SARIMA')
plt.legend(loc='best')
plt.show()

lylayang

from statsmodels.tsa.stattools import adfuller
df1=df.resample('D', how=np.mean)

def test_stationarity(timeseries):
    rolmean = timeseries.rolling(window=30).mean()
    rolstd = timeseries.rolling(window=30).std()
    
    plt.figure(figsize=(14,5))
    sns.despine(left=True)
    orig = plt.plot(timeseries, color='blue',label='Original')

lylayang

train_score = model.evaluate(trainX, trainY, verbose=0)
print('Train Root Mean Squared Error(RMSE): %.2f; Train Mean Absolute Error(MAE) : %.2f ' 
% (np.sqrt(train_score[1]), train_score[2]))
test_score = model.evaluate(testX, testY, verbose=0)
print('Test Root Mean Squared Error(RMSE): %.2f; Test Mean Absolute Error(MAE) : %.2f ' 
% (np.sqrt(test_score[1]), test_score[2]))
model_loss(history)

lylayang

def convert2matrix(data_arr, look_back):
 X, Y =[], []
 for i in range(len(data_arr)-look_back):
  d=i+look_back  
  X.append(data_arr[i:d,0])
  Y.append(data_arr[d,0])
 return np.array(X), np.array(Y)

lylayang

#Split data set into testing dataset and train dataset
train_size = 900
train, test =df1.values[0:train_size,:],df1.values[train_size:len(df1.values),:]
# setup look_back window 
look_back = 30
#convert dataset into right shape in order to input into the DNN
trainX, trainY = convert2matrix(train, look_back)
testX, testY = convert2matrix(test, look_back)

lylayang

train_size = 900
train,test = df1.values[0:train_size,:], df1.values[train_size:len(df1.values),:]
look_back = 30 #create window size as look_back=30
test = np.append(test,np.repeat(test[-1,], look_back))
train = np.append(train,np.repeat(train[-1,],look_back))
trainX,trainY =convert2matrix(train,look_back)
testX,testY =convert2matrix(test, look_back)
# reshape input to be [samples, window size, features]
trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))
testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))

lylayang

train_size = 900
test_size = len(df_arr) - train_size
train, test = df_arr[0:train_size,:], df_arr[train_size:len(df_arr),:]
look_back = 30
trainX, trainY = convert2matrix(train, look_back)
testX, testY = convert2matrix(test, look_back)
# reshape input to be [samples, time steps, features]
trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))
testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))

lylayang

from sklearn.preprocessing import MinMaxScaler
#create numpy.ndarray
df_arr= df1.values 
df1.values = df_arr.astype('float32')
df_arr = np.reshape(df_arr, (-1, 1)) #LTSM requires more input features compared to RNN or DNN
scaler = MinMaxScaler(feature_range=(0, 1))#LTSM is senstive to the scale of features
df_arr = scaler.fit_transform(df_arr)

lylayang

train_predict = model.predict(trainX)
test_predict = model.predict(testX)
# invert predictions
train_predict = scaler.inverse_transform(train_predict)
trainY = scaler.inverse_transform([trainY])
test_predict = scaler.inverse_transform(test_predict)
testY = scaler.inverse_transform([testY])
print('Train Root Mean Squared Error(RMSE): %.2f; Train Mean Absolute Error(MAE) : %.2f '% (np.sqrt(mean_squared_error(trainY[0], train_predict[:,0])),(mean_absolute_error(trainY[0], train_predict[:,0]))))
print('Test Root Mean Squared Error(RMSE): %.2f; Test Mean Absolute Error(MAE) : %.2f '% (np.sqrt(mean_squared_error(testY[0], test_predict[:,0])),(mean_absolute_error(testY[0], test_predict[:,0]))))
model_loss(history)