train.py

import pandas as pd
import numpy as np
from pandas_datareader import data
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler
from sklearn.neural_network import MLPClassifier
from sklearn.metrics import classification_report,confusion_matrix
from datetime import datetime

def stat(X_test, predictions, y_test, position_sizing, accuracy):
    X_start = X_test.ix[:, 0]
    Trade_test = (y_test / accuracy - X_start)
    Trade_pred = (predictions / accuracy - X_start)
    Trade_diff = pd.concat([Trade_test, Trade_pred, X_start, y_test/accuracy], axis=1)
    summ = 0
    summ_n = 0
    trend = 0
    trend_false = 0
    for row in Trade_diff.iterrows():
        trend += 1
        if row[1].values[0] > 0 and row[1].values[1] > 0:
            summ += (row[1].values[0] / row[1].values[2]) * position_sizing / 100
        elif row[1].values[0] < 0 and row[1].values[1] < 0:
            summ -= (row[1].values[0] / row[1].values[2]) * position_sizing / 100
        else:
            summ -= (row[1].values[0].__abs__() / row[1].values[2]) * position_sizing / 100
            summ_n += (row[1].values[0].__abs__() / row[1].values[2]) * position_sizing / 100
            trend_false += 1
    print("count {}, average absolute {}% , profit {}% , loss {}% , trend failing {}%".format(
        len(Trade_diff),
        round((predictions - y_test).abs().div(predictions).sum() / predictions.size * 100, 2),
        round(summ*100, 2), round(summ_n*100, 2), round(trend_false/trend*100, 2)))

#stock = data.DataReader('SBER', 'moex', start='2004-01-01', end='2018-09-30')
#stock_TQBR = stock[stock.BOARDID == 'TQBR']
#X = stock_TQBR['CLOSE'].dropna()
#SBER_040101_181002
# http://export.finam.ru/MGNT_080101_181002.csv?market=1&em=17086&code=MGNT&apply=0&df=1&mf=0&yf=2008&from=01.01.2008&dt=2&mt=9&yt=2018&to=09.10.2018&p=8&f=MGNT_080101_181002&e=.csv&cn=MGNT&dtf=1&tmf=1&MSOR=1&mstime=on&mstimever=1&sep=1&sep2=1&datf=5&at=1
X = pd.read_csv(
    '~/Downloads/MGNT_080101_181009.csv',
    parse_dates=[0], index_col=0
)['<CLOSE>'].astype(float)

accuracy = 10
prediction_days = 10
window_days = 200
test_tail_days = 0
predic_tail_days = 5
X_window = pd.concat(
    [X.shift(w) for w in range(0, window_days, 1)], axis=1
)
# Add day of week
#X_dayofweek = X.index.dayofweek.astype(int)
#for d in range(5):
#    X_window.insert(loc=d+1, column='dayofweek{}'.format(d), value=np.where(X_dayofweek==d, 1, 0))

y = X.shift(-prediction_days)*accuracy
X_train, X_test, y_train, y_test = train_test_split(
    X_window.iloc[window_days-1:-prediction_days-test_tail_days],
    y.iloc[window_days-1:-prediction_days-test_tail_days].astype(int),
    test_size = 0.15,
    shuffle=True
)
X_test = X_test.append(X_window.iloc[-prediction_days-test_tail_days:-prediction_days])
y_test = y_test.append(y.iloc[-prediction_days-test_tail_days:-prediction_days].astype(int))
X_predic = X_window[-predic_tail_days:]
# Fit only to the training data
# Now apply the transformations to the data:
scaler = MinMaxScaler()
scaler.fit(X.values.reshape(-1,1))
X_train_norm = scaler.transform(X_train)
#Reverse
#X_train_norm = scaler.transform(X_train.reindex(index=X_train.index[::-1]))
#y = y.reindex(index=y.index[::-1])
X_test_norm = scaler.transform(X_test)
X_predic_norm = scaler.transform(X_predic)
#  train our model
mlp = MLPClassifier(activation='relu', solver='lbfgs', max_iter=1100, alpha=0.015, hidden_layer_sizes=(375),
                    random_state=1)
mlp.fiat(X_train_norm, y_train)
predictions = mlp.predict(X_test_norm)
new_predictions = mlp.predict(X_predic_norm)
stat(X_test, predictions, y_test, 25, accuracy)

#print((predictions-y_test).sort_index().tail()/100)
#print((predictions-y_test).div(predictions).abs().mul(100) .describe())
#print((predictions - y_test).abs().div(predictions).sum() / predictions.size * 100)
# mlp = MLPClassifier(activation='relu', solver='lbfgs',max_iter=1000, alpha=0.0001, hidden_layer_sizes=(200)
# 2.7549654896139275
# logistic
# total % 273.2222664780387, negativ % 144.6650856196263, trend % 30.252100840336134
# relu test_size 0.5 alpha=0.015 hidden_layer_sizes=375 max_iter=1100
# count 596, average absolute 1.83% , profit 707.8% , loss 13.86% , trend failing 8.56%
# count 598, average absolute 1.88% , profit 723.24% , loss 22.78% , trend failing 10.87%
# relu test_size 0.75 alpha=0.015 hidden_layer_sizes = 375 max_iter 1000
# count 894, average absolute 2.38% , profit 1004.97% , loss 66.52% , trend failing 12.98%
# relu test_size 0.75 alpha=0.015 hidden_layer_sizes = 375 max_iter 1100
# count 894, average absolute 2.11% , profit 996.95% , loss 60.53% , trend failing 13.31%
# w160 count 862, average absolute 2.17% , profit 964.68% , loss 57.2% , trend failing 14.15%
# w180 count 848, average absolute 2.2% , profit 969.73% , loss 46.88% , trend failing 12.74%
# w220 count 820, average absolute 2.64% , profit 916.49% , loss 76.91% , trend failing 15.24%
# 25 days
# count 883, average absolute 3.15% , profit 1700.48% , loss 86.18% , trend failing 11.1%
# 20 days random
# count 886, average absolute 2.47% , profit 1627.53% , loss 26.85% , trend failing 8.47%
# count 886, average absolute 2.65% , profit 1528.17% , loss 50.05% , trend failing 10.72%
# count 591, average absolute 1.94% , profit 1045.88% , loss 21.53% , trend failing 8.97%
# 15 days
# count 594, average absolute 1.8% , profit 365.24% , loss 10.05% , trend failing 10.1%
# count 594, average absolute 1.95% , profit 395.5% , loss 4.95% , trend failing 5.89%
# 10 days
# count 894, average absolute 2.46% , profit 970.95% , loss 81.8% , trend failing 14.99%
# count 596, average absolute 1.75% , profit 281.79% , loss 7.06% , trend failing 9.9%
# count 597, average absolute 1.73% , profit 287.0% , loss 8.33% , trend failing 12.06%
# dayofweek count 596, average absolute 4.07% , profit 642.56% , loss 77.19% , trend failing 17.79%
# dayofweek count 596, average absolute 4.38% , profit 582.84% , loss 88.09% , trend failing 20.13%
# 5 days random
# count 838, average absolute 2.3% , profit 522.07% , loss 91.74% , trend failing 21.72%
# count 838, average absolute 2.09% , profit 539.57% , loss 84.15% , trend failing 19.21%
# w250 count 803, average absolute 2.47% , profit 438.98% , loss 121.09% , trend failing 24.41%
# w180 count 852, average absolute 2.29% , profit 558.84% , loss 81.27% , trend failing 20.07%

# MGNT
# count 1243, average absolute 2.31% , profit 1580.98% , loss 56.97% , trend failing 11.91%
# count 621, average absolute 1.62% , profit 764.6% , loss 20.59% , trend failing 10.31%