今天來利用RNN的LSTM訓練模型，來預測股票吧!

我們以元大50 2022年1-12月的收盤價作為訓練資料。

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from sklearn.preprocessing import MinMaxScaler
from keras.models import Sequential
from keras.layers import LSTM, Dense
from plotly.graph_objs import Scatter, Layout
from plotly.offline import plot

定義train與test資料

def load_data(df, sequence_length=10, split=0.8):  #載入時間序列資料
    data_all = np.array(df).astype(float)    # 轉為浮點數據
    scaler = MinMaxScaler() 
    data_all = scaler.fit_transform(data_all)  # 將數據縮放為 0~1 之間
    data = []
    
    for i in range(len(data_all) - sequence_length - 1):
        # 每筆 data 資料有 11 欄
        data.append(data_all[i: i + sequence_length + 1])
        
    reshaped_data = np.array(data).astype('float64')
    x = reshaped_data[:, :-1] # 第 1至第10個欄位為 特徵
    y = reshaped_data[:, -1]  # 第 11個欄位為 label
    
    split_boundary = int(reshaped_data.shape[0] * split)
    train_x = x[: split_boundary] # 前 80% 為 train 的特徵
    test_x = x[split_boundary:]   # 最後 20% 為 test 的特徵
 
    train_y = y[: split_boundary] # 前 80% 為 train 的 label
    test_y = y[split_boundary:]   # 最後 20% 為 test 的 label

    return train_x, train_y, test_x, test_y, scaler

建立一個循環神經網絡

def build_model():
    model = Sequential()     
    # 隱藏層：256 個神經元，input_shape：(10,1)
    # TIME_STEPS=10,INPUT_SIZE=1
    model.add(LSTM(input_shape=(10,1),units=256,unroll=False))
    model.add(Dense(units=1)) # 輸出層：1 個神經元
    #compile:loss, optimizer, metrics
    model.compile(loss="mse", optimizer="adam", metrics=['accuracy'])
    return model

進行預測

def train_model(train_x, train_y, test_x, test_y):
    model = build_model()
    try:
        model.fit(train_x, train_y, batch_size=100, epochs=300, validation_split=0.1)
        predict = model.predict(test_x)
        predict = np.reshape(predict, (predict.size, )) #轉換為1維矩陣
    except KeyboardInterrupt:
        print(predict)
        print(test_y)
    return predict, test_y #傳回 預測值和真實值

讀取檔案

pd.options.mode.chained_assignment = None  #取消顯示pandas資料重設警告
filename = 'twstockyear2022.csv'
df = pd.read_csv(filename, encoding='big5')  #以pandas讀取檔案
ddprice=pd.DataFrame(df['收盤價'])

模型訓練和預測，並還原預測結果和真實資料的數值範圍

train_x, train_y, test_x, test_y, scaler =load_data(ddprice, sequence_length=10, split=0.8)

predict_y, test_y = train_model(train_x, train_y, test_x, test_y)
predict_y = scaler.inverse_transform([[i] for i in predict_y]) # 還原
test_y = scaler.inverse_transform(test_y)  # 還原

將預測結果和真實資料可視化為一個散點圖

plt.plot(predict_y, 'y:') #預測
plt.plot(test_y, 'g-')    #收盤價
plt.legend(['predict', 'price'])
plt.show()

dd2=pd.DataFrame({"predict":list(predict_y),"label":list(test_y)})
#轉換為 numpy 陣列，並轉為 float
dd2["predict"] = np.array(dd2["predict"]).astype('float64')
dd2["label"] = np.array(dd2["label"]).astype('float64')

data = [
    Scatter(y=dd2["predict"],name='預測'),
    Scatter(y=dd2["label"],name='收盤價')
] 

plot({"data": data, "layout": Layout(title='2022年個股預測圖')},auto_open=True)

結果

---20231010---