經過研究，發現RSRS的正確用法其實是需要用到兩個數據，分別是
n: 一階擬合樣本數，m:求均值方差樣本數，其中n比較小 如18，m比較大 如1100
經過調優后，收益率顯著上升！ 如下圖：
（并且算法上優化，不重復計算）
?將源碼貼出來，感興趣的點個贊和關注，謝謝！
# 導入函數庫
from jqdata import *
import pandas as pd
import numpy as np
import statsmodels.api as smdef initialize(context):set_benchmark('000300.XSHG')set_option('use_real_price', True)set_order_cost(OrderCost(close_tax=0.001, open_commission=0.0003, close_commission=0.0003, min_commission=5), type='stock')g.security = '000300.XSHG'g.buy_beta, g.sell_beta = 0.7, -0.7g.is_first, g.beta_list, g.r2_list = True, [], []run_daily(market_open, time='open', reference_security='000300.XSHG')def calculate_rsrs(end_date='', n=18, m=1100):# n: 一階擬合樣本數，m:求均值方差樣本數; 采用普通最小二乘法擬合，前面算過的就不重復計算了if g.is_first:df = get_price(g.security, end_date=end_date, count=m+n, frequency='daily', fields=['high','low'])else:df = get_price(g.security, end_date=end_date, count=n, frequency='daily', fields=['high','low'])for i in range(len(df))[(-m if g.is_first else -1):]:x = sm.add_constant(df['low'][i-n+1:i+1])y = df['high'][i-n+1:i+1]model = sm.OLS(y, x).fit()beta = model.params[1]r2 = model.rsquaredg.beta_list.append(beta)g.r2_list.append(r2) section = g.beta_list[-m:]mu = np.mean(section)sigma = np.std(section)z_score = (section[-1] - mu)/sigmaz_score_right = z_score * beta * r2g.is_first = Falsereturn z_score_rightdef market_open(context):cash = context.portfolio.available_cashz_score_right = calculate_rsrs(context.previous_date)if z_score_right>g.buy_beta and cash>0:order_value(g.security, cash)elif z_score_right<g.sell_beta and context.portfolio.positions[g.security].closeable_amount > 0:order_target(g.security, 0)?
?
(C/C++))
