Merge branch 'master' of github.com:wassname/rl-portfolio-management

readme.md

@@ -1,8 +1,8 @@
 Attempting to replicate "A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem" by [Jiang et. al. 2017](https://arxiv.org/abs/1706.10059) [1].
 
-**note the authors have put [the official code for the paper up](https://github.com/ZhengyaoJiang/PGPortfolio)**
+**Note: the papers authors have put [the official code for the paper up and it works well](https://github.com/ZhengyaoJiang/PGPortfolio)**
 
-tl;dr I managed to get 8% growth on training data, but it disapeared on test data. However, RL papers can be very difficult to replicate due to bugs, framework differences, and hyperparameter sensistivity.
+tl;dr I managed to get 8% growth on training data, but it disapeared on test data. However, RL papers can be very difficult to replicate due to bugs, framework differences, and hyperparameter sensistivity. 
 
 # About
 

requirements/requirements.txt

@@ -3,7 +3,7 @@
 gym==0.9.3
 tensorflow-gpu==1.3.0
 h5py==2.7.0
-# tensorforce 0.3.5.2 a specific commit
+# tensorforce 0.3.5.1 a specific commit
 https://github.com/reinforceio/tensorforce/archive/4b5741d2088869e7f8f40c354bd63762c8d62833.zip
 # numbers
 tables==3.4.2

rl_portfolio_management/util.py

@@ -6,12 +6,17 @@ def sharpe(returns, freq=30, rfr=0.0):
     return (np.sqrt(freq) * np.mean(returns - rfr)) / (np.std(returns - rfr) + eps)
 
 
-def MDD(returns):
-    """Max drawdown."""
-    peak = returns.max()
-    i = returns.argmax()
-    trough = returns[i:].min()
-    return (trough - peak) / trough
+def MDD(X):
+    """By nicktids, see issue 15."""
+    mdd = 0
+    peak = X[0]
+    for x in X:
+        if x > peak: 
+            peak = x
+        dd = (peak - x) / peak
+        if dd > mdd:
+            mdd = dd
+    return mdd
 
 
 def softmax(w, t=1.0):