Merge branch 'master' of https://github.com/Nico-Chaves/EE378_Code

expert_pool.py

@@ -23,7 +23,7 @@ class ExpertPool(Portfolio):
     def __init__(self, market_data, experts, start=0, stop=None, init_weights=None,
                  rebal_interval=1, tune_interval=None,
                  init_b=None, init_dollars=init_dollars,
-                 weighting_strategy='exp_window', windows=[10], ew_alpha=0.5, ew_eta=0.1,
+                 weighting_strategy='exp_window', windows=[10], ew_alpha=0.5, ew_eta=0.8,
                  verbose=False, silent=False, past_results_dir=None, new_results_dir=None, repeat_past=False):
 
         if not isinstance(market_data, MarketData):

olmar.py

@@ -13,10 +13,11 @@ class OLMAR(Portfolio):
     Online Moving Average Reversion (OLMAR) Portfolio
 
     Introduced by Li and Hoi. "On-Line Portfolio Selection with Moving Average Reversion"
+    http://icml.cc/2012/papers/168.pdf
 
     """
-    def __init__(self, market_data, market_data_train=None, start=0, stop=None, window=20, eps=1.1, rebal_interval=1,
-                 window_range=range(16, 26, 2), eps_range=[1.1, 1.2, 1.3, 1.4, 1.5], tune_interval=None,
+    def __init__(self, market_data, market_data_train=None, start=0, stop=None, window=10, eps=1.3, rebal_interval=1,
+                 window_range=range(5, 30, 3), eps_range=np.arange(1.1, 5.1, 0.2), tune_interval=15,
                  init_b=None, verbose=False, silent=False, past_results_dir=None, new_results_dir=None, repeat_past=False):
         """
 
@@ -138,7 +139,6 @@ def get_new_allocation(self, day, init=False):
 
         # limit lambda to avoid numerical problems from acting too aggressively.
         # (referenced from marigold's implementation: https://github.com/Marigold/universal-portfolios)
-        # TODO: check if this is necessary
         lam = min(100000, lam)
 
         # Note: we don't perform simplex project b/c negative values (shorting) is allowed.
@@ -156,8 +156,8 @@ def tune_hyperparams(self, cur_day):
         # Create new instances of this portfolio with various hyperparameter settings
         # to find the best constant hyperparameters in hindsight
 
-        tune_duration = 20
-        if cur_day >= tune_duration:
+        tune_duration = 10  # Tune over the last 2 weeks
+        if cur_day > tune_duration:
             start_day = cur_day - tune_duration
         else:
             # Not worth tuning yet
@@ -166,8 +166,7 @@ def tune_hyperparams(self, cur_day):
         hyperparam_space = [self.window_range, self.eps_range]
         hyp_combos = list(itertools.product(*hyperparam_space))
 
-        #init_b = self.b_history[-tune_duration]  # Allocation used at beginning of tuning period
-        init_b = None
+        init_b = self.b_history[:,cur_day-tune_duration]   # Allocation used at beginning of tuning period
 
         # Compute sharpe ratios for each setting of hyperparams
         sharpe_ratios = []

portfolio.py

@@ -62,7 +62,7 @@ def __init__(self, market_data, market_data_train=None, start=0, stop=None, reba
             past_b_history, past_dollars_history = self.load_previous_results(past_results_dir)
             self.past_b_history = past_b_history
             self.past_dollars_history = past_dollars_history
-            self.len_past = past_b_history.shape[1]
+            self.len_past = past_b_history.shape[1] - 1
             self.b = past_b_history[:, -1]  # Use previous b as initialization (overrides |init_b| argument)
         else:
             self.past_b_history = None
@@ -85,7 +85,7 @@ def update(self, cur_day, init=False):
 
         # Check if we need to tune hyperparameters today
         if self.tune_interval and not self.repeat_past:
-            if cur_day > self.start and cur_day % self.tune_interval == 0:
+            if cur_day > self.start and cur_day % self.tune_interval == 1:
                 self.tune_hyperparams(cur_day)
 
         self.update_allocation(cur_day, init)
@@ -212,7 +212,8 @@ def save_results(self):
         save_dir = self.new_results_dir
 
         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)
+        full_b = np.concatenate((self.b_history, self.b.reshape(-1, 1)), axis=1)
+        util.save_b_history(save_dir=save_dir, b_history=full_b, portfolio_type=self.portfolio_type)
         util.save_hyperparams(save_dir=save_dir, hyperparams_dict=self.get_hyperparams_dict(), portfolio_type=self.portfolio_type)
         return
 

rmr.py

@@ -16,9 +16,9 @@ class RMR(OLMAR):
     http://www.ijcai.org/Proceedings/13/Papers/296.pdf
 
     """
-    def __init__(self, market_data, market_data_train=None, start=0, stop=None, window=20, eps=1.3, tau=0.01, max_iter=100,
-                rebal_interval=1, window_range=range(14, 26, 2), eps_range=[1.1, 1.3, 1.5, 1.7, 1.9, 2.1, 2.3, 2.5],
-                 tune_interval=None, init_b=None, verbose=False, silent=False,
+    def __init__(self, market_data, market_data_train=None, start=0, stop=None, window=20, eps=1.5, tau=0.001, max_iter=100,
+                rebal_interval=1, window_range=range(5, 30, 3), eps_range=np.arange(1.1, 5.1, 0.2),
+                 tune_interval=25, init_b=None, verbose=False, silent=False,
                  past_results_dir=None, new_results_dir=None, repeat_past=False):
 
         self.portfolio_type = 'RMR'
@@ -99,12 +99,7 @@ def predict_price_relatives(self, day):
             mu_avail_full_window = self.T_func(mu_avail_full_window, window_pr_avail_full_window)
             L1_dist = np.linalg.norm((prev_mu-mu_avail_full_window), ord=1)
             thresh = self.tau * np.linalg.norm(mu_avail_full_window, ord=1)
-            """
-            prev_mu = mu
-            mu = self.T_func(mu, window_pr_avail_today)
-            L1_dist = np.linalg.norm((prev_mu-mu), ord=1)
-            thresh = self.tau * np.linalg.norm(mu, ord=1)
-            """
+
             if L1_dist <= thresh:
                 break
 
@@ -156,22 +151,13 @@ def T_func(self, mu, window_cl):
 
         T_tilde = s2 * 1.0 / s1
         return T_tilde
-        """
-        gamma = np.linalg.norm(R_tilde, ord=2)
-        gamma_inv = 1.0 / gamma
-
-        return T_bar + min(1.0, gamma_inv) * mu
-        """
-
 
     def tune_hyperparams(self, cur_day):
         # Create new instances of this portfolio with various hyperparameter settings
         # to find the best constant hyperparameters in hindsight
 
-        # TODO: use previous history, if available
-
-        tune_duration = 20
-        if cur_day >= tune_duration:
+        tune_duration = 10  # Tune over the last 2 weeks
+        if cur_day > tune_duration:
             start_day = cur_day - tune_duration
         else:
             # Not worth tuning yet
@@ -180,8 +166,7 @@ def tune_hyperparams(self, cur_day):
         hyperparam_space = [self.window_range, self.eps_range]
         hyp_combos = list(itertools.product(*hyperparam_space))
 
-        #init_b = self.b_history[-tune_duration]  # Allocation used at beginning of tuning period
-        init_b = None
+        init_b = self.b_history[:,cur_day-tune_duration]   # Allocation used at beginning of tuning period
 
         # Compute sharpe ratios for each setting of hyperparams
         sharpe_ratios = []