第24节 比特币vs黄金,莱特币vs白银

abupy/FactorBuyBu/ABuFactorBuyDemo.py

@@ -7,6 +7,9 @@
 from __future__ import print_function
 from __future__ import division
 
+import numpy as np
+import pandas as pd
+
 from .ABuFactorBuyBase import AbuFactorBuyBase, BuyCallMixin
 from .ABuFactorBuyBreak import AbuFactorBuyBreak
 from ..TLineBu.ABuTL import AbuTLine
@@ -206,3 +209,129 @@ def make_ump_block_decision(self, ml_feature_dict):
  if result == -1:
  return True
  return False
+
+
+class AbuBTCDayBuy(AbuFactorBuyBase, BuyCallMixin):
+ """
+ 比特币日交易策略：
+
+ 1. 买入条件1: 当日这100个股票60%以上都是上涨的
+ 2. 买入条件2: 使用在第12节：机器学习与比特币示例中编写的：信号发出今天比特币会有大行情
+ """
+
+ def _init_self(self, **kwargs):
+ from ..UtilBu.ABuProgress import AbuProgress
+ from ..MarketBu import ABuSymbolPd
+
+ # 市场中与btc最相关的top个股票
+ self.btc_similar_top = kwargs.pop('btc_similar_top')
+ # 超过多少个相关股票今天趋势相同就买入
+ self.btc_vote_val = kwargs.pop('btc_vote_val', 0.60)
+ self.pg = AbuProgress(len(self.kl_pd), 0, 'btc buy day')
+
+ def _collect_kl(sim_line):
+ """在初始化中将所有相关股票的对应时间的k线数据进行收集"""
+ start = self.kl_pd.iloc[0].date
+ end = self.kl_pd.iloc[-1].date
+ kl = ABuSymbolPd.make_kl_df(sim_line.symbol, start=start, end=end)
+ self.kl_dict[sim_line.symbol] = kl
+
+ self.kl_dict = {}
+ # k线数据进行收集到类字典对象self.kl_dict中
+ self.btc_similar_top.apply(_collect_kl, axis=1)
+
+ # noinspection PyUnresolvedReferences
+ def fit_day(self, today):
+ """
+ :param today: 当前驱动的交易日金融时间序列数据
+ :return:
+ """
+ self.pg.show()
+
+ # key是金融时间序列索引
+ day_ind = int(today.key)
+ # 忽略不符合买入的天（统计周期内前两天及最后一天），因为btc的机器学习特证需要三天交易数据
+ if day_ind < 2 or day_ind >= self.kl_pd.shape[0] - 1:
+ return None
+
+ # 今天，昨天，前天三天的交易数据进行特证转换
+ btc = self.kl_pd[day_ind - 2:day_ind + 1]
+ # 三天的交易数据进行转换后得到btc_today_x
+ btc_today_x = self.make_btc_today(btc)
+
+ # btc_ml并没有在这里传入，实际如果要使用，需要对外部的btc_ml进行本地序列化后，构造读取本地
+ # 买入条件2: 使用在第12节：机器学习与比特币示例中编写的：信号发出今天比特币会有大行情
+ if btc_ml.predict(btc_today_x):
+ # 买入条件1: 当日这100个股票60%以上都是上涨的
+ vote_val = self.similar_predict(today.date)
+ if vote_val > self.btc_vote_val:
+ return self.make_buy_order(day_ind - 1)
+
+ # noinspection PyUnresolvedReferences
+ def make_btc_today(self, sib_btc):
+ """构造比特币三天数据特证"""
+ from ..UtilBu import ABuScalerUtil
+
+ sib_btc['big_wave'] = (sib_btc.high - sib_btc.low) / sib_btc.pre_close > 0.55
+ sib_btc['big_wave'] = sib_btc['big_wave'].astype(int)
+ sib_btc_scale = ABuScalerUtil.scaler_std(
+ sib_btc.filter(['open', 'close', 'high', 'low', 'volume', 'pre_close',
+ 'ma5', 'ma10', 'ma21', 'ma60', 'atr21', 'atr14']))
+ # 把标准化后的和big_wave，date_week连接起来
+ sib_btc_scale = pd.concat([sib_btc['big_wave'], sib_btc_scale, sib_btc['date_week']], axis=1)
+
+ # 抽取第一天，第二天的大多数特征分别改名字以one，two为特征前缀，如：one_open，one_close，two_ma5，two_high.....
+ a0 = sib_btc_scale.iloc[0].filter(['open', 'close', 'high', 'low', 'volume', 'pre_close',
+ 'ma5', 'ma10', 'ma21', 'ma60', 'atr21', 'atr14', 'date_week'])
+ a0.rename(index={'open': 'one_open', 'close': 'one_close', 'high': 'one_high', 'low': 'one_low',
+ 'volume': 'one_volume', 'pre_close': 'one_pre_close',
+ 'ma5': 'one_ma5', 'ma10': 'one_ma10', 'ma21': 'one_ma21',
+ 'ma60': 'one_ma60', 'atr21': 'one_atr21', 'atr14': 'one_atr14',
+ 'date_week': 'one_date_week'}, inplace=True)
+
+ a1 = sib_btc_scale.iloc[1].filter(['open', 'close', 'high', 'low', 'volume', 'pre_close',
+ 'ma5', 'ma10', 'ma21', 'ma60', 'atr21', 'atr14', 'date_week'])
+ a1.rename(index={'open': 'two_open', 'close': 'two_close', 'high': 'two_high', 'low': 'two_low',
+ 'volume': 'two_volume', 'pre_close': 'two_pre_close',
+ 'ma5': 'two_ma5', 'ma10': 'two_ma10', 'ma21': 'two_ma21',
+ 'ma60': 'two_ma60', 'atr21': 'two_atr21', 'atr14': 'two_atr14',
+ 'date_week': 'two_date_week'}, inplace=True)
+ # 第三天的特征只使用'open', 'low', 'pre_close', 'date_week'，该名前缀today，如today_open，today_date_week
+ a2 = sib_btc_scale.iloc[2].filter(['big_wave', 'open', 'low', 'pre_close', 'date_week'])
+ a2.rename(index={'open': 'today_open', 'low': 'today_low',
+ 'pre_close': 'today_pre_close',
+ 'date_week': 'today_date_week'}, inplace=True)
+ # 将抽取改名字后的特征连接起来组合成为一条新数据，即3天的交易数据特征－>1条新的数据
+ btc_today = pd.DataFrame(pd.concat([a0, a1, a2], axis=0)).T
+
+ # 开始将周几进行离散处理
+ dummies_week_col = btc_ml.df.filter(regex='(^one_date_week_|^two_date_week_|^today_date_week_)').columns
+ dummies_week_df = pd.DataFrame(np.zeros((1, len(dummies_week_col))), columns=dummies_week_col)
+
+ # 手动修改每一天的one hot
+ one_day_key = 'one_date_week_{}'.format(btc_today.one_date_week.values[0])
+ dummies_week_df[one_day_key] = 1
+ two_day_key = 'two_date_week_{}'.format(btc_today.two_date_week.values[0])
+ dummies_week_df[two_day_key] = 1
+ today_day_key = 'today_date_week_{}'.format(btc_today.today_date_week.values[0])
+ dummies_week_df[today_day_key] = 1
+ btc_today.drop(['one_date_week', 'two_date_week', 'today_date_week'], inplace=True, axis=1)
+ btc_today = pd.concat([btc_today, dummies_week_df], axis=1)
+ return btc_today.as_matrix()[:, 1:]
+
+ def similar_predict(self, today_date):
+ """与比特币在市场中最相关的top100个股票已各自今天的涨跌结果进行投票"""
+ def _predict_vote(sim_line, _today_date):
+ kl = self.kl_dict[sim_line.symbol]
+ if kl is None:
+ return -1 * sim_line.vote_direction > 0
+ kl_today = kl[kl.date == _today_date]
+ if kl_today is None or kl_today.empty:
+ return -1 * sim_line.vote_direction > 0
+ # 需要 * sim_line.vote_direction，因为负相关的存在
+ return kl_today.p_change.values[0] * sim_line.vote_direction > 0
+
+ vote_result = self.btc_similar_top.apply(_predict_vote, axis=1, args={today_date, })
+ # 将所有投票结果进行统计，得到与比特币最相关的这top100个股票的今天投票结果
+ vote_val = 1 - vote_result.value_counts()[False] / vote_result.value_counts().sum()
+ return vote_val

abupy/FactorSellBu/ABuFactorSellNDay.py

@@ -0,0 +1,41 @@
+# -*- encoding:utf-8 -*-
+"""
+ 卖出择时示例因子：n日卖出策略，不管什么结果，买入后只持有N天
+"""
+
+from __future__ import absolute_import
+from __future__ import print_function
+from __future__ import division
+
+from .ABuFactorSellBase import AbuFactorSellBase, filter_sell_order, skip_last_day, ESupportDirection
+
+__author__ = '阿布'
+__weixin__ = 'abu_quant'
+
+
+class AbuFactorSellNDay(AbuFactorSellBase):
+ """n日卖出策略，不管交易现在什么结果，买入后只持有N天"""
+
+ def _init_self(self, **kwargs):
+ """kwargs中可以包含: 参数sell_n：代表买入后持有的天数，默认1天"""
+ self.sell_n = kwargs.pop('sell_n', 1)
+ self.sell_type_extra = '{}:sell_n={}'.format(self.__class__.__name__, self.sell_n)
+
+ def support_direction(self):
+ """因子支持两个方向"""
+ return [ESupportDirection.DIRECTION_CAll.value, ESupportDirection.DIRECTION_PUT.value]
+
+ @skip_last_day
+ @filter_sell_order
+ def fit_day(self, today, orders):
+ """
+ :param today: 当前驱动的交易日金融时间序列数据
+ :param orders: 买入择时策略中生成的订单序列
+ :return:
+ """
+ for order in orders:
+ # 将单子的持有天数进行增加
+ order.keep_days += 1
+ if order.keep_days >= self.sell_n:
+ # 只要超过self.sell_n即卖出
+ order.fit_sell_order(int(today.key), self)

abupy/FactorSellBu/__init__.py

@@ -5,6 +5,7 @@
 from .ABuFactorAtrNStop import AbuFactorAtrNStop
 from .ABuFactorCloseAtrNStop import AbuFactorCloseAtrNStop
 from .ABuFactorSellBreak import AbuFactorSellBreak
+from .ABuFactorSellNDay import AbuFactorSellNDay
 # noinspection all
 from . import ABuFS as fs
 
@@ -17,6 +18,7 @@
  'AbuFactorAtrNStop',
  'AbuFactorCloseAtrNStop',
  'AbuFactorSellBreak',
+ 'AbuFactorSellNDay',
 
  'fs'
 ]

abupy/MLBu/ABuMLPd.py

@@ -158,20 +158,34 @@ class BtcBigWaveClf(AbuMLPd):
  做出这个判断的依据里有一条即是今天的波动需要足够大
  """
 
- def make_xy(self, **kwarg):
-
- # 从沙盒中读取测试数据
- ABuEnv.enable_example_env_ipython()
- btc = ABuSymbolPd.make_kl_df('btc', start='2013-09-01', end='2017-07-26')
- ABuEnv.disable_example_env_ipython()
+ def __init__(self, **kwarg):
+ """
+ 如果传递了btc数据，说明不是运行：
+ 12-机器学习与比特币示例(ABU量化使用文档)
+ """
+ self.btc = kwarg.pop('btc', None)
+ super(BtcBigWaveClf, self).__init__(**kwarg)
 
+ def make_xy(self, **kwarg):
+ if self.btc is None:
+ # 从沙盒中读取测试数据
+ ABuEnv.enable_example_env_ipython()
+ btc = ABuSymbolPd.make_kl_df('btc', start='2013-09-01', end='2017-07-26')
+ ABuEnv.disable_example_env_ipython()
+ else:
+ btc = self.btc
  # .055的日震荡幅度可以成做大波动的交易对比特币来说，下面对数据添加新列big_wave
  btc['big_wave'] = (btc.high - btc.low) / btc.pre_close > 0.055
  btc['big_wave'] = btc['big_wave'].astype(int)
 
- # 首先切割训练集和测试集，保留最后60天走势数据做为测试集数据
- btc_train_raw = btc[:-60]
- btc_test_raw = btc[-60:]
+ if self.btc is None:
+ # 如果是12-机器学习与比特币示例(ABU量化使用文档)，保留60天数据
+ # 首先切割训练集和测试集，保留最后60天走势数据做为测试集数据
+ btc_train_raw = btc[:-60]
+ btc_test_raw = btc[-60:]
+ else:
+ btc_train_raw = btc
+ btc_test_raw = None
 
  # 下面为训练集和测试集数据都加上5，10，21，60日均线特征
  def calc_ma(tc, p_ma):
@@ -180,7 +194,8 @@ def calc_ma(tc, p_ma):
 
  for ma in [5, 10, 21, 60]:
  calc_ma(btc_train_raw, ma)
- calc_ma(btc_test_raw, ma)
+ if btc_test_raw is not None:
+ calc_ma(btc_test_raw, ma)
  # 下面使用训练集数据btc_train_raw做为参数抽取组合特征，重新组合好的特征
  btc_train0 = self.btc_siblings_df(btc_train_raw)
  # 由于每3条连续交易日数据组合成一个特征，只要向前跳一条数据进行特征组合抽取即可以得到另一组新特征
@@ -194,29 +209,29 @@ def calc_ma(tc, p_ma):
  dummies_two_week = pd.get_dummies(btc_train['two_date_week'], prefix='two_date_week')
  dummies_today_week = pd.get_dummies(btc_train['today_date_week'], prefix='today_date_week')
  btc_train.drop(['one_date_week', 'two_date_week', 'today_date_week'], inplace=True, axis=1)
-
  # make_xy中需要确定self.df
  self.df = pd.concat([btc_train, dummies_one_week, dummies_two_week, dummies_today_week], axis=1)
  # make_xy中需要确定x, y
- train_matrix = btc_train.as_matrix()
+ train_matrix = self.df.as_matrix()
  self.y = train_matrix[:, 0]
  self.x = train_matrix[:, 1:]
 
- # 下面将前面保留切割的60条测试数据进行特征抽取组合，方式和抽取训练集时一样
- btc_test0 = self.btc_siblings_df(btc_test_raw)
- btc_test1 = self.btc_siblings_df(btc_test_raw[1:])
- btc_test2 = self.btc_siblings_df(btc_test_raw[2:])
- btc_test = pd.concat([btc_test0, btc_test1, btc_test2])
- btc_test.index = np.arange(0, btc_test.shape[0])
- dummies_one_week = pd.get_dummies(btc_test['one_date_week'], prefix='one_date_week')
- dummies_two_week = pd.get_dummies(btc_test['two_date_week'], prefix='two_date_week')
- dummies_today_week = pd.get_dummies(btc_test['today_date_week'], prefix='today_date_week')
- btc_test.drop(['one_date_week', 'two_date_week', 'today_date_week'], inplace=True, axis=1)
- self.btc_test = pd.concat([btc_test, dummies_one_week, dummies_two_week, dummies_today_week], axis=1)
- # 测试集数据构建
- matrix_test = btc_test.as_matrix()
- self.y_test = matrix_test[:, 0]
- self.x_test = matrix_test[:, 1:]
+ if btc_test_raw is not None:
+ # 下面将前面保留切割的60条测试数据进行特征抽取组合，方式和抽取训练集时一样
+ btc_test0 = self.btc_siblings_df(btc_test_raw)
+ btc_test1 = self.btc_siblings_df(btc_test_raw[1:])
+ btc_test2 = self.btc_siblings_df(btc_test_raw[2:])
+ btc_test = pd.concat([btc_test0, btc_test1, btc_test2])
+ btc_test.index = np.arange(0, btc_test.shape[0])
+ dummies_one_week = pd.get_dummies(btc_test['one_date_week'], prefix='one_date_week')
+ dummies_two_week = pd.get_dummies(btc_test['two_date_week'], prefix='two_date_week')
+ dummies_today_week = pd.get_dummies(btc_test['today_date_week'], prefix='today_date_week')
+ btc_test.drop(['one_date_week', 'two_date_week', 'today_date_week'], inplace=True, axis=1)
+ self.btc_test = pd.concat([btc_test, dummies_one_week, dummies_two_week, dummies_today_week], axis=1)
+ # 测试集数据构建
+ matrix_test = self.btc_test.as_matrix()
+ self.y_test = matrix_test[:, 0]
+ self.x_test = matrix_test[:, 1:]
 
  # noinspection PyMethodMayBeStatic
  def btc_siblings_df(self, btc_raw):

abupy/RomDataBu/futures_gb.csv

@@ -0,0 +1,20 @@
+,symbol,product,min_deposit,min_unit,exchange
+0,NID,伦敦镍,0.07,6,LME
+1,PBD,伦敦铅,0.1,25,LME
+2,SND,伦敦锡,0.05,5,LME
+3,ZSD,伦敦锌,0.1,25,LME
+4,AHD,伦敦铝,0.08,25,LME
+5,CAD,伦敦铜,0.08,25,LME
+6,XAU,伦敦金,0.1,1,LME
+7,XAG,伦敦银,0.1,100,LME
+8,XPT,伦敦铂金,0.1,1,LME
+9,S,美黄豆,0.1,5000,CBOT
+10,W,美小麦,0.1,5000,CBOT
+11,C,美玉米,0.1,5000,CBOT
+12,BO,美豆油,0.1,60000,CBOT
+13,SM,美豆粕,0.1,100,CBOT
+14,HG,纽约铜,0.1,25000,NYMEX
+15,SI,纽约白银,0.08,5000,NYMEX
+16,GC,纽约黄金,0.08,100,NYMEX
+17,CL,纽约原油,0.07,1000,NYMEX
+18,NG,纽约天然气,0.07,10000,NYMEX

abupy/TradeBu/ABuOrder.py

@@ -44,7 +44,7 @@ class AbuOrder(object):
  # 多个因子买入条件可能生成几百万个order对象使用__slots__降低内存消耗
  __slots__ = ('order_deal', 'buy_symbol', 'buy_date', 'buy_factor', 'buy_price', 'buy_cnt', 'buy_pos', 'sell_date',
  'buy_type_str', 'expect_direction',
- 'sell_type', 'sell_price', 'sell_type_extra', 'ml_features')
+ 'sell_type', 'keep_days', 'sell_price', 'sell_type_extra', 'ml_features')
 
  def __init__(self):
  """初始设置只需要将order_deal设置未成交状态"""
@@ -162,6 +162,9 @@ def fit_buy_order(self, day_ind, factor_object):
  self.sell_date = None
  # 订单卖出类型，keep：持有
  self.sell_type = 'keep'
+ # 交易日持有天数
+ self.keep_days = 0
+
  # 订单卖出价格
  self.sell_price = None
  # 订单卖出额外信息