BackTestOptimization.py

BackTestOptimization.py

# -*- coding: utf-8 -*-
import poloniex
import time
import datetime
from sklearn import tree
import numpy as np
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
import pickle


def main():
    # --トレーニング用パラメータ
    #トレーニングデータ数
    theNumberOfTrainData_Min=20
    theNumberOfTrainData_Max = 40
    # トレーニングデータと教師データのセット数
    theNumberOfTrainAndKyoushiSet_Min=20
    theNumberOfTrainAndKyoushiSet_Max = 40

    studyTrialTimes = 200 #予測のトライアル数(予測結果がばらつくので1回の予測結果を出すために実施する試行回数(取りあえずこのままにしておいてください。)

    # --バックテストパラメータ
    plotGraphFlag = 0  # バックテスト結果の資金グラフをプロットしますか？yes:1。no:0 (最適化時実行は0にしてください。)
    initialFund = 1000  # 初期資金
    spread = 0  # 取引スプレット(手数料)
    backTestDays = 30*2 # バックテスト期間(日)

    dateBTC, dataBTC = getDataPoloniex()
    dateBTC.reverse()
    dataBTC.reverse()
    data = changeData(dataBTC)


    X = np.arange(theNumberOfTrainData_Min,theNumberOfTrainData_Max+1,1)
    Y = np.arange(theNumberOfTrainAndKyoushiSet_Min,theNumberOfTrainAndKyoushiSet_Max+1,1)
    X, Y = np.meshgrid(X, Y)
    Z = np.zeros([len(Y[:]), len(X[0])])

    for i in range(0, len(X[0])):
        for j in range(0, len(Y[:])):
            theNumberOfTrainData=X[j][i]
            theNumberOfTrainAndKyoushiSet=Y[j][i]
            seitouritsuUp, seitouritsuDown, increasedFundRatio, increasedBTCPriceRatio, simOutputStr = backTest(data, dateBTC, dataBTC, theNumberOfTrainData,theNumberOfTrainAndKyoushiSet, studyTrialTimes, initialFund, spread, backTestDays, plotGraphFlag)
            Z[j][i]=increasedFundRatio
            print(simOutputStr)
            print('---------------------------------------')

    maxZRow = np.where(Z == np.max(Z))[0][0]
    maxZCol = np.where(Z == np.max(Z))[1][0]

    theNumberOfTrainData_Opt=X[maxZRow][maxZCol]
    theNumberOfTrainAndKyoushiSet_Opt=Y[maxZRow][maxZCol]
    theDateOpt=datetime.datetime.now()

    backTestOptResult={"X":X,"Y":Y,"Z":Z,"theNumberOfTrainData_Opt":theNumberOfTrainData_Opt,"theNumberOfTrainAndKyoushiSet_Opt":theNumberOfTrainAndKyoushiSet_Opt,"theDateOpt":theDateOpt}
    with open('backTestOptResult.pickle', mode='wb') as f:
        pickle.dump(backTestOptResult, f)

    print('最適解(資金上昇率): ' + str(round(Z[maxZRow][maxZCol]*100,1)) + '%')
    print('最適トレーニングデータ数: ' + str(theNumberOfTrainData_Opt))
    print('最適トレーニングデータと教師データのセット数: ' + str(theNumberOfTrainAndKyoushiSet_Opt))

    fig = plt.figure()
    ax = Axes3D(fig)
    ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=plt.cm.hot)
    ax.contourf(X, Y, Z, zdir='z', offset=-2, cmap=plt.cm.hot)
    ax.set_xlabel('theNumberOfTrainData')
    ax.set_ylabel('theNumberOfTrainAndKyoushiSet')
    ax.set_zlabel('increasedFundRatio')
    plt.show()


def backTest(data, dateBTC, dataBTC, theNumberOfTrainData,theNumberOfTrainAndKyoushiSet, studyTrialTimes, initialFund, spread, backTestDays, plotGraphFlag):
    realValue = []
    predictionValue = []
    fund = [initialFund]
    pastDay = 0
    seitouUp = 0
    seitouDown = 0
    for trainStartDay in range(backTestDays, 0, -1):
        y_Prediction, y = predictionTommorowBTCForBackTest(data, trainStartDay, theNumberOfTrainData, theNumberOfTrainAndKyoushiSet, studyTrialTimes)
        realValue.append(y)
        predictionValue.append(y_Prediction)
        pastDay += 1
        if y_Prediction == y:
            if y_Prediction == 1:
                seitouUp += 1
                fund.append(fund[pastDay - 1] * (1 + abs(data[trainStartDay - 1]) - spread))
            else:
                seitouDown += 1
                #---fund.append(fund[pastDay - 1] * (1 + abs(data[trainStartDay - 1]) - spread))
                fund.append(fund[pastDay - 1])
        else:
            if y_Prediction == 1:
                fund.append(fund[pastDay - 1] * (1 - abs(data[trainStartDay - 1]) - spread))
            else:
                #---fund.append(fund[pastDay - 1] * (1 - abs(data[trainStartDay - 1]) - spread))
                fund.append(fund[pastDay - 1])

    # ----バックテスト結果出力
    outputStr = '・トレーニングデータ数: ' + str(theNumberOfTrainData) +'\n'
    outputStr += '・トレーニングデータと教師データのセット数: ' + str(theNumberOfTrainAndKyoushiSet)+'\n'
    outputStr += '・バックテスト期間: 過去' + str(backTestDays) + '日\n'
    seitouritsuUp = float(seitouUp) / sum(predictionValue)
    seitouritsuDown = float(seitouDown) / (backTestDays - sum(predictionValue))
    outputStr += '・正答率(上昇時): ' + str(round(seitouritsuUp * 100,1)) + '%' + '\n'
    outputStr += '・正答率(下降時): ' + str(round(seitouritsuDown * 100,1)) + '%' + '\n'
    trainStartDay=0
    dataBTCOnBackTest = dataBTC[trainStartDay:trainStartDay+backTestDays+1]
    dataBTCOnBackTest.reverse()
    increasedFundRatio=(fund[-1]-fund[0])/fund[0]
    increasedBTCPriceRatio=(dataBTCOnBackTest[-1]-dataBTCOnBackTest[0])/dataBTCOnBackTest[0]
    outputStr += '・資金上昇率: ' + str(round(increasedFundRatio*100,1)) + '% (初期資金は$' +str(fund[0])+'。最終資金は$' +str(fund[-1])+'。)' + '\n'
    outputStr += '・BTC価格上昇率: ' + str(round(increasedBTCPriceRatio*100,1)) + '% (初期BTC価格は$' +str(dataBTCOnBackTest[0])+'。最終BTC価格は$' +str(dataBTCOnBackTest[-1])+'。)' + '\n'

    # ----資金推移グラフ
    if(plotGraphFlag):
        # ----Plot Fund
        dateBTCOnBackTest = dateBTC[trainStartDay:trainStartDay + backTestDays + 1]
        dateBTCOnBackTest.reverse()
        fig1, ax1 = plt.subplots()
        p1, = ax1.plot(dateBTCOnBackTest, fund, '-ob')
        ax1.set_title("Simulation with past data")
        ax1.set_xlabel("Day")
        ax1.set_ylabel("Fund[$]")
        plt.grid(fig1)
        # ----Plot BTC Price
        ax2 = ax1.twinx()
        p2, = ax2.plot(dateBTCOnBackTest, dataBTCOnBackTest, '-or')
        ax2.set_ylabel('BTC price[$]')
        ax1.legend([p1, p2], ["Fund", "BTC price"], loc="upper left")
        plt.show(fig1)
    return seitouritsuUp, seitouritsuDown, increasedFundRatio, increasedBTCPriceRatio, outputStr

def getDataPoloniex():
    polo = poloniex.Poloniex()
    polo.timeout = 2
    chartUSDT_BTC = polo.returnChartData('USDT_BTC', period=polo.DAY, start=time.time() - polo.DAY * 500, end=time.time())
    tmpDate = [chartUSDT_BTC[i]['date'] for i in range(len(chartUSDT_BTC))]
    date = [datetime.datetime.fromtimestamp(tmpDate[i]).date() for i in range(len(tmpDate))]
    data = [float(chartUSDT_BTC[i]['open']) for i in range(len(chartUSDT_BTC))]
    return date ,data

def changeData(data):
    newData=[]
    for i in range(0, len(data) - 1):
        newData.append(float(data[i] - data[i + 1]) / data[i + 1])
    newData.append(0)
    return newData

def preparationTrainAndKyoushiSets(data,trainStartDay,theNumberOfTrainData,theNumberOfTrainAndKyoushiSet):
    train_X = []
    train_y = []
    for i in range(0,theNumberOfTrainAndKyoushiSet):
        train_X.append(data[trainStartDay+1+i:trainStartDay+theNumberOfTrainData+1+i])
        train_y.append([int(0 < data[trainStartDay+i])])
    return train_X, train_y

def predictionTommorowBTCForBackTest(data, trainStartDay, theNumberOfTrainData,theNumberOfTrainAndKyoushiSet, studyTrialTimes):
    train_X, train_y = preparationTrainAndKyoushiSets(data, trainStartDay, theNumberOfTrainData,theNumberOfTrainAndKyoushiSet)
    X=data[trainStartDay:trainStartDay+theNumberOfTrainData]
    y = int(0 < data[trainStartDay-1])
    y_PredictionTommorow=[]
    for i in range(0, theNumberOfTrainAndKyoushiSet):
        clf = tree.DecisionTreeClassifier()
        clf.fit(train_X, train_y)
        y_PredictionTommorow.append(clf.predict([X])[0])
    upOrDownRatio = sum(y_PredictionTommorow) * 1.0 / len(y_PredictionTommorow)
    return int(upOrDownRatio>0.5),y


if __name__ == "__main__":
    main()