-
Notifications
You must be signed in to change notification settings - Fork 5
/
nonparametric_markowitz.py
204 lines (162 loc) · 8.64 KB
/
nonparametric_markowitz.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
import numpy as np
from constants import init_dollars
import util
from portfolio import Portfolio
#import matplotlib.pyplot as plt
from cvxpy import *
import pdb
class NonParametricMarkowitz(Portfolio):
def __init__(self, market_data, market_data_train=None, window_len=10, k=10, risk_aversion=1e-5, start_date=25, start=0, stop=None,
rebal_interval=1, tune_interval=None, tune_length=None,
init_b=None, init_dollars=init_dollars, verbose=False, silent=False,
past_results_dir=None, new_results_dir=None, repeat_past=False):
self.portfolio_type = 'NPM'
self.window_len = window_len
self.k = k
self.risk_aversion = risk_aversion
self.start_date = start_date
self.mu = None
self.sigma = None
self.startup_time = 10
if past_results_dir is not None:
hyperparams_dict = util.load_hyperparams(past_results_dir, ['Window', 'K', 'Risk', 'Start_Date'])
self.window_len = int(hyperparams_dict['Window'])
self.k = int(hyperparams_dict['K'])
self.risk_aversion = hyperparams_dict['Risk']
self.start_date = hyperparams_dict['Start_Date']
self.load_state(past_results_dir)
super(NonParametricMarkowitz,self).__init__(market_data, market_data_train=market_data_train, start=start, stop=stop,
rebal_interval=rebal_interval, tune_interval=tune_interval, tune_length=tune_length,
init_b=None, init_dollars=init_dollars, verbose=False, silent=False, past_results_dir=past_results_dir,
new_results_dir=new_results_dir, repeat_past=repeat_past)
def get_market_window(self, window, day):
# Compose historical market window, including opening prices
available = util.get_avail_stocks(self.data.get_op()[day - window + 1, :])
if(day <= window-1):
available = util.get_avail_stocks(self.data_train.get_op()[-1, :])
available_inds = np.asarray([i for i in range(497) if available[i] > 0])
if(day >= window-1):
op = self.data.get_op()[day-window+1:day+1,available_inds]
hi = self.data.get_hi()[day-window+1:day,available_inds]
lo = self.data.get_lo()[day-window+1:day,available_inds]
cl = self.data.get_cl()[day-window+1:day,available_inds]
elif(self.data_train is not None):
op = self.data_train.get_op()[day-window+1:, available_inds]
op = np.concatenate((op,self.data.get_op()[:day+1,available_inds]))
hi = self.data_train.get_hi()[day-window+1:, available_inds]
hi = np.concatenate((hi,self.data.get_hi()[:day,available_inds]))
lo = self.data_train.get_lo()[day-window+1:, available_inds]
lo = np.concatenate((lo,self.data.get_lo()[:day,available_inds]))
cl = self.data_train.get_cl()[day-window+1:, available_inds]
cl = np.concatenate((cl,self.data.get_cl()[:day,available_inds]))
else:
raise 'NPM called get_market_window with day<window'
history = np.concatenate((op, hi, lo, cl)).T
return history
def get_new_allocation(self, cur_day, init=False):
"""
:param cur_day:
:return: A (1 x num_stocks) array of fractions. Each fraction represents the
amount of the money should be invested in that stock at the end of the day.
If we haven't reached the start date (the day when we have confidence in our mean/covariance info), return
a uniform portfolio. Otherwise, perform nonparametric markowitz.
"""
self.update_statistics(cur_day)
if cur_day == 0:
cur_day_op = self.data.get_op(relative=False)[cur_day, :] # opening prices on |cur_day|
return util.get_uniform_allocation(self.num_stocks, cur_day_op)
elif(cur_day < self.start_date):
available = util.get_avail_stocks(self.data.get_op()[cur_day, :])
num_available = np.sum(available)
new_allocation = 1.0/num_available * np.asarray(available)
else:
k = self.k
available = util.get_avail_stocks(self.data.get_op()[cur_day - self.window_len + 1, :])
if(cur_day <= self.window_len-1):
available = util.get_avail_stocks(self.data_train.get_op()[-1, :])
available_inds = util.get_available_inds(available)
history = self.get_market_window(self.window_len, cur_day)
num_available = history.shape[0]
neighbors = np.zeros((num_available, k))
history_norms = np.diag(np.dot(history, history.T))
# Compute k nearest neighbors for each stock
for i in range(num_available):
stock = history[i,:]
neighbors[i,:] = util.k_nearest_neighbors(stock, history, k, history_norms)
# Solve optimization problem
l = self.risk_aversion
neighbors = neighbors.astype(int)
b = Variable(num_available)
c = Variable(num_available)
d = 0
for i in range(num_available):
inds = available_inds[neighbors[i,:]]
m_i = self.mu[inds] #np.ones(inds.shape[0])
S_i = self.sigma[inds,:]
S_i = S_i[:,inds]
d += -(b[i]*m_i).T*np.ones(k) + l*quad_form(b[i]*np.ones(k), S_i) #+ 0.00005*norm(b-b_last,2) #0.00005
constraints = [c>= b, c >= -b, sum_entries(c)==1] #, b <= self.cap, b>= - self.cap] #[b >= 0, np.ones(num_available).T*b == 1]
objective = Minimize(d)
prob = Problem(objective, constraints)
prob.solve()
new_allocation = np.zeros(len(available))
new_allocation[available_inds] = b.value
if(cur_day%50 == 0):
print 'Day %d' % (cur_day)
return new_allocation
def update_statistics(self, cur_day):
'''
Perform mean and covariance estimation.
Currently, we assume independence (diagonal sigma) for the first start_date days, then move to the true ML estimate,
hoping that our covariance does not become singular later on.
'''
# Get the previous day's closing data
last_close = np.nan_to_num(self.data.get_cl()[cur_day-1,:])
num_total_stocks = last_close.shape[0]
num_examples = cur_day-1
if(self.data_train is not None):
num_examples += 1247
# If the parameters are uninitialized, initialize them
if(self.mu == None):
assert(self.sigma == None)
self.mu = np.zeros(num_total_stocks).astype(float)
self.sigma = np.zeros((num_total_stocks,num_total_stocks)).astype(float)
if(cur_day > 1):
N = num_examples-1
mu_N = self.mu
self.mu = N/(N+1.0) * self.mu + 1/(N+1.0) * last_close
if(cur_day >= self.start_date):
self.sigma = N/(N+1.0) * (self.sigma+np.outer(mu_N, mu_N)) - np.outer(self.mu, self.mu) + 1/(N+1.0)*np.outer(last_close, last_close)
else:
self.sigma = N/(N+1.0) * (self.sigma + np.diag(mu_N**2)) - np.diag(self.mu**2) + 1/(N+1.0)*np.diag(last_close**2)
def get_hyperparams_dict(self):
hyperparams = {
'Window': str(self.window_len),
'K': str(self.k),
'Risk': str(self.risk_aversion),
'Start_Date' : str(self.start_date),
}
return hyperparams
def save_results(self):
if self.new_results_dir is None:
return
print 'Saving ', self.portfolio_type
save_dir = self.new_results_dir
# Add final portfolio
self.b_history = np.concatenate((self.b_history, np.reshape(self.b, (-1,1))), axis=1)
util.save_dollars_history(save_dir=save_dir, dollars=self.dollars_op_history, portfolio_type=self.portfolio_type)
util.save_b_history(save_dir=save_dir, b_history=self.b_history, portfolio_type=self.portfolio_type)
util.save_hyperparams(save_dir=save_dir, hyperparams_dict=self.get_hyperparams_dict(), portfolio_type=self.portfolio_type)
self.save_state(save_dir=save_dir)
return
def save_state(self, save_dir):
np.save(save_dir + 'mu.npy', self.mu)
np.save(save_dir + 'sigma.npy', self.sigma)
def load_state(self, load_dir):
self.mu = np.load(load_dir + 'mu.npy')
self.sigma = np.load(load_dir + 'sigma.npy')
def print_results(self):
print 30 * '-'
print 'Performance for Nonparametric Markowitz Portfolio:'
print 30 * '-'
Portfolio.print_results(self)