超簡単Pythonで株価予測（keras・RNN 利用）ディープラーニング

Pythonでkerasを利用して翌日の株価の上下予測を超簡単にディープラーニング（RNN使用）

1. ツールインストール

$ pip install scikit-learn keras pandas-datareader

2. ファイル作成

pred.py

import numpy as np
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
from keras import Sequential
from keras.layers import Dense, SimpleRNN
import pandas_datareader as pdr
from sklearn.metrics import accuracy_score

df = pdr.get_data_yahoo("AAPL", "2010-11-01", "2020-11-01")
df["Diff"] = df.Close.diff()
df["SMA_2"] = df.Close.rolling(2).mean()
df["Force_Index"] = df.Close * df.Volume
df["y"] = df["Diff"].apply(lambda x: 1 if x > 0 else 0).shift(-1)
df = df.drop(
   ["Open", "High", "Low", "Close", "Volume", "Diff", "Adj Close"],
   axis=1,
).dropna()
# print(df)
X = StandardScaler().fit_transform(df.drop(["y"], axis=1))
y = df["y"].values
X_train, X_test, y_train, y_test = train_test_split(
   X,
   y,
   test_size=0.2,
   shuffle=False,
)
model = Sequential()
model.add(SimpleRNN(2, input_shape=(X_train.shape[1], 1)))
model.add(Dense(1, activation="sigmoid"))
model.compile(optimizer="adam", loss="binary_crossentropy", metrics=["acc"])
model.fit(X_train[:, :, np.newaxis], y_train, epochs=100)
y_pred = model.predict(X_test[:, :, np.newaxis])
print(accuracy_score(y_test, y_pred > 0.5))

3. 実行

$ python pred.py

Epoch 1/100
63/63 [==============================] - 0s 1ms/step - loss: 0.6959 - acc: 0.4950
Epoch 2/100
63/63 [==============================] - 0s 2ms/step - loss: 0.6947 - acc: 0.4980
:
Epoch 99/100
63/63 [==============================] - 0s 1ms/step - loss: 0.6920 - acc: 0.5224
Epoch 100/100
63/63 [==============================] - 0s 1ms/step - loss: 0.6921 - acc: 0.5194

0.5376984126984127

以上、超簡単！

4. 結果

同じデータ、特徴量で、計算した結果、XGBoost・DNN・LSTM・GRU・RNN・LogisticRegression・k-nearest neighbor・RandomForest・BernoulliNB・SVM・RGF・MLP・Bagging・Voting・Stacking・LightGBM・TCN・HGBCのうちMLPが最も良いという事に

XGBoost            0.5119047619047619
DNN                0.5496031746031746
LSTM               0.5178571428571429
GRU                0.5138888888888888
RNN                0.5376984126984127
LogisticRegression 0.5496031746031746
k-nearest neighbor 0.5198412698412699
RandomForest       0.49603174603174605
BernoulliNB        0.5496031746031746
SVM                0.5396825396825397
RGF                0.5158730158730159
MLP                0.5694444444444444
Bagging            0.5297619047619048
Voting             0.5416666666666666
Stacking           0.5218253968253969
LightGBM           0.5456349206349206
TCN                0.5198412698412699
HGBC               0.5

5. 参考

Keras documentation: SimpleRNN layer