【学习笔记】-使用LSTM算法实现余额宝资金流入流出预测

使用LSTM算法实现余额宝资金流入流出预测

关键词：LSTM、基于大规模历史数据预测、MSE
数据来源：[天池大赛-资金流入流出预测-挑战Baseline]
数据预处理：根据数据集进行数据预处理生成每日购入资金总量，最终传入的数据列表如下

模型代码运行过程：
RNN模型运行数据

预测数据与原始数据对比图：

MSE、RMSE、MAE结果：

LSTM模型完整代码：
（数据量较大，代码运行时间较长，需要耐心等待）

#以下部分为项目需要引用的库
from math import sqrt
import numpy
import pandas
from keras.layers import LSTM
from keras.layers import Dense
from keras.callbacks import EarlyStopping
import matplotlib.pyplot as plt
from sklearn.metrics import mean_squared_error, mean_absolute_error
from random import choice
from sklearn.preprocessing import MinMaxScaler
from keras.models import Sequential
import warnings
warnings.filterwarnings('ignore')


#设置学习循环的次数（epoch）神经元个数（neure）步长（seq_len）的范围
numpy.random.seed(2019)
epoch_range=[*range(100,200,1)]
neure_range=[*range(10, 128, 1)]
seq_len_range=[*range(5, 50, 1)]

#定义参数组的list
List=[]
combination_all=[]
combination_best=[]
pred_best=[]
pred = []

epoch_best=None
neure_best=None
seq_len_best=None
MSE_best=399044410397529222

def Change_num(index_X,X_all_list):
    if(3<=index_X<=(len(X_all_list)-3)):
        list_X=X_all_list[index_X-3:index_X+2]
    elif(index_X<3):
        list_X=X_all_list[:5]
    else:
        list_X=X_all_list[-5:]
    return list_X
    
#随机生成一组参数值
for i in range(0,1):
    epoch=choice(epoch_range)
    neure=choice(neure_range)
    seq_len=choice(seq_len_range)
    List.append(1)
    List[i]=[epoch,neure,seq_len]
print(List)

leean=2

#RNN模型主体
class RNNModel(object):
    def __init__(self, look_back=1, epochs_purchase=10, batch_size=1, verbose=2, patience=10,
                 store_result=False,nerue=1):
        self.look_back = look_back
        self.epochs_purchase = epochs_purchase
        self.batch_size = batch_size
        self.verbose = verbose
        self.store_result = store_result
        self.patience = patience
        self.nerue = nerue
        self.purchase = pandas.read_csv('/Users/Desktop/Purchase Redemption Data/total_balance_LSTM.csv', usecols=[1], engine='python')


#数据标准化
    def get_Standard_data(self, data_frame):
        # load the data set
        data_set= data_frame.values
        data_set = data_set.astype('float32')
        scaler = MinMaxScaler(feature_range=(0, 1))
        data_set = scaler.fit_transform(data_set)
        train_x, train_y, test = self.create_data_set(data_set)
        train_x = numpy.reshape(train_x, (train_x.shape[0], 1, train_x.shape[1]))

        return train_x, train_y, test, scaler

    def create_data_set(self, data_set):
        data_x, data_y = [], []
        print(self.look_back)
        for i in range(len(data_set) - self.look_back - 31):
            a = data_set[i:(i + self.look_back), 0]
            data_x.append(a)
            data_y.append(list(data_set[i + self.look_back: i + self.look_back + 31, 0]))
        return numpy.array(data_x), numpy.array(data_y), data_set[-self.look_back:, 0].reshape(1, 1, self.look_back)

#LSTM模型主体
    def LSTM_model(self, train_x, train_y, epochs, nerue):
        model = Sequential()
        model.add(LSTM(nerue, input_shape=(1, self.look_back),return_sequences=True))
        model.add(LSTM(32, return_sequences=False))
        model.add(Dense(32))
        model.add(Dense(31))
        model.compile(loss='mse', optimizer='adam')
        model.summary()
        early_stopping = EarlyStopping('loss', patience=self.patience)
        history = model.fit(train_x, train_y, epochs=epochs, batch_size=self.batch_size, verbose=self.verbose,
                            callbacks=[early_stopping])
        return model

    def predict(self, model, data):
        prediction = model.predict(data)
        return prediction

#处理流入数据，划分训练集和测试集，并运行LSTM算法`在这里插入代码片`
    def run_purchase(self):
        purchase_train_x, purchase_train_y, purchase_test, purchase_scaler = self.get_Standard_data(self.purchase)
        purchase_model = self.LSTM_model(purchase_train_x, purchase_train_y, self.epochs_purchase,self.nerue)
        purchase_predict = self.predict(purchase_model, purchase_test)
        purchase = purchase_scaler.inverse_transform(purchase_predict).reshape(31, 1)
        test_user = pandas.DataFrame({'report_date': [20140800 + i for i in range(1, 32)]})
        test_user['purchase'] = purchase
        return test_user['purchase']


if __name__ == '__main__':
    real = pandas.read_csv('/Users/Desktop/Purchase Redemption Data/total_balance_test.csv',usecols=[1], engine='python')

#抽取27组epoch、neure、seq_len的参数组合
    for i in range(0, 27):
        epoch = choice(epoch_range)
        neure = choice(neure_range)
        seq_len = choice(seq_len_range)
        List.append(1)
        List[i] = [epoch, neure, seq_len]
    print(List)

    leean = 2
    for f in range(3):
        MSE_list = []
        pred_all = []
        for s in range(27):
            initiation = RNNModel(look_back=List[s][2], epochs_purchase=List[s][0], batch_size=14, verbose=0,
                                  patience=50,
                                  store_result=True,nerue=List[s][1])
            prediction=initiation.run_purchase()
            MSE_list.append(1)
            MSE_list[s] = mean_squared_error(real, prediction)
            print(MSE_list[s])
            pred_all.append(1)
            pred_all[s] = prediction
        print(MSE_list)
        indexs = MSE_list.index(min(MSE_list))
        print(indexs)
        pred_best.append(1)
        pred_best[f] = pred_all[indexs]
        epochs, neures, seq_lens = List[indexs][0], List[indexs][1], List[indexs][2]
        combination_best.append(1)
        combination_best[f] = [epochs, neures, seq_lens]
#计算不同参数的预测模型的预测结果的MSE值，根据MSE值选择最合适的参数组合
        if (MSE_list[indexs] < MSE_best):
            MSE_best = MSE_list[indexs]
            epoch_best = epochs
            neure_best = neures
            seq_len_best = seq_lens
            pred = pred_all[indexs]
            print(pred[0])
        index_epoch = epoch_range.index(epochs)
        del epoch_range[index_epoch]
        index_nerue = neure_range.index(neures)
        del neure_range[index_nerue]
        index_seq_len = seq_len_range.index(seq_lens)
        del seq_len_range[index_seq_len]
        epoch_new = Change_num(index_epoch, epoch_range)
        neure_new = Change_num(index_nerue, neure_range)
        seq_len_new = Change_num(index_seq_len, seq_len_range)

        List = []
        for y in range(27):
            List.append(1)
            List[y] = [choice(epoch_new), choice(neure_new), choice(seq_len_new)]
        combination_all.append(1)
        combination_all[f] = List
        print(f, "th selection！\n")
        print("epoch:", epoch_best, "nerue:", neure_best, "seq_len:", seq_len_best, "\n")
        print("Minimize MSE:", MSE_best)

#输出预测结果最好的的最佳参数组合
    print("All of the combination of the parameters", combination_all)
    print("The best combination of the parameters=", combination_best)
    print("The best combination of the parameters：\n")
    print("epoch:", epoch_best, "nerue:", neure_best, "seq_len:", seq_len_best, "\n")
    print("MSE:", MSE_best)

#绘制曲线图
    plt.figure(figsize=(6, 4))
    plt.plot(real, label='Actual')
    plt.plot(pred, color='red', label='Predicted')
    plt.xticks()
    plt.title('Actual vs Predicted LSTM_purchase')
    plt.xlabel('date')
    plt.ylabel('purchase')
    X=[0,4,9,14,19,24,30]
    Y=['2014-08-01','2014-08-05','2014-08-10','2014-08-15','2014-08-20','2014-08-25','2014-08-31']
    plt.xticks(X,Y)
    plt.legend()
    plt.show()
    
#输出预测结果的MSE、RMSE、MAE值
    mseValue = mean_squared_error(real, pred)
    rmseValue = sqrt(mean_squared_error(real, pred))
    maeValue = mean_absolute_error(real, pred)
    print('MSE: %.6f' % mseValue)
    print('RMSE: %.6f' % rmseValue)
    print('MAE: %.6f' % maeValue)
    ```