lin_regress.py

import numpy as np
import pandas as pd
from pandas_datareader import data as web
import seaborn as sns
import matplotlib.pyplot as plt
import datetime as dt
sns.set()

class RegPred(object):
    def __init__(self, symbol):
        self.symbol = symbol
        self.get_data()

    def get_data(self):
        self.data = pd.DataFrame(web.DataReader(self.symbol, data_source='yahoo',start= dt.datetime(2015,1,1),end= dt.datetime(2017,1,26))['Adj Close'])
        self.data.sort_index(ascending=False,inplace=True)
        self.data.columns = ['prices']
        self.data['returns'] = np.log(self.data / self.data.shift(-1))
        self.data.dropna(inplace=True)

    def generate_matrix(self, lags):
        self.matrix = np.zeros((lags + 1, len(self.data) - lags))
        for i in range(lags + 1):
            if i == lags:
                self.matrix[i] = self.data.returns.values[i:]
            else:
                self.matrix[i] = self.data.returns.values[i: i - lags]

    def predict_returns(self, lags):
        self.lags = lags
        self.generate_matrix(lags)
        # reg = np.linalg.lstsq(self.matrix[:lags].T, np.sign(self.matrix[lags]))[0]
        reg = np.linalg.lstsq(self.matrix[1:].T, self.matrix[0])[0]

        self.pred = np.dot(self.matrix[:lags].T, reg)

    def get_performance(self):
        self.perf = self.data.ix[:-(self.lags+1)].copy()
        self.perf['positions'] = np.sign(self.pred[1:]) # 0th record is the next day prediction prediction where we havent got the data yet
        self.perf['strategy'] = self.perf.positions * self.perf.returns
        self.perf[['returns', 'strategy']].cumsum().apply(np.exp).plot()

# class NNPred(object):

if __name__ == "__main__":
    lag = 6 #<-- keep changing the lags!!
    regClass = RegPred("MSFT")
    regClass.generate_matrix(lag)
    regClass.predict_returns(lag)
    regClass.get_performance()
    plt.show()