【统计套利】配对交易

来源：https://uqer.io/community/share/559c85baf9f06cb5614f190d

策略思路

寻找走势相关且股价相近的一对股票，根据其价格变动买卖

策略实现

• 首先，历史前五日的Pearson相关系数若大于给定的阈值
• 如果两只股票走势趋同，则按上涨（下跌）趋势买入（卖出）股票
• 如果两只股票走势背离，则买入下跌股票，卖出上涨股票

首先需要在A股中寻找走势相关性很大的股票，这是一项很繁复的工作。为简单起见，这里直接使用了一个现成的结果：000159和000967在2012年的走势十分相似，这一点可以通过复权收盘价曲线来验证

<matplotlib.axes.AxesSubplot at 0x6415790>

接下来我们对2013年全年的这两只股票进行配对交易策略的回测：

from scipy.stats.stats import pearsonr

start = '2013-01-01'
end  = '2014-01-01'
benchmark = 'HS300'
universe = ['000159.XSHE', '000967.XSHE']
capital_base = 1000000

def initialize(account):
    account.cutoff = 0.9
    account.prev_prc1 = 0
    account.prev_prc2 = 0
    account.prev_prcb = 0

def handle_data(account):
    if len(account.universe) < 2: return

    clsp = account.get_attribute_history('closePrice', 5)
    stk1, stk2 = universe
    px1, px2 = clsp[stk1], clsp[stk2]

    prc1, prc2 = px1[-1], px2[-1]
    prcb = account.get_symbol_history('benchmark', 1)['return'][0]

    if account.prev_prc1 == 0:
        account.prev_prc1 = prc1
        account.prev_prc2 = prc2
        account.prev_prcb = prcb
        return

    corval, pval = pearsonr(px1, px2)

    if abs(corval) < account.cutoff:
        return

    mov1, mov2 = adj(prc1, prc2, prcb, account.prev_prc1, account.prev_prc2, account.prev_prcb)

    amount = 100000 / prc2    
    if mov1 > 0:
        order(stk2, amount)
    elif mov1 < 0:
        if account.valid_secpos.get(stk2, 0) > amount:
            order(stk2, -amount)
        else:
            order_to(stk2, 0)

    amount = 100000 / prc1
    if mov2 > 0:
        order(stk1, amount)
    elif mov2 < 0:
        if account.valid_secpos.get(stk1, 0) > amount:
            order(stk1, -amount)
        else:
            order_to(stk1, 0)

    account.prev_prc1 = prc1
    account.prev_prc2 = prc2
    account.prev_prcb = prcb

def adj(x, y, base, prev_x, prev_y, prev_base):
    dhs = base / prev_base - 1
    dx = x / prev_x - 1 - dhs
    dy = y / prev_y - 1 - dhs
    return dx, dy

bt

233 rows × 6 columns

根据bt的表现，在仓位控制和组合管理方面应该还有着不少进一步优化的空间