poliniex to poloniex

0. load poliniex data 30m multindex.ipynb

@@ -4,7 +4,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "load 30m poliniex data"
+    "load 30m poloniex data"
    ]
   },
   {
@@ -52,7 +52,7 @@
    "outputs": [],
    "source": [
     "dfs=[]\n",
-    "for infile in glob.glob('./data/poliniex_teachmehowtotrade/*.csv'):\n",
+    "for infile in glob.glob('./data/poloniex_teachmehowtotrade/*.csv'):\n",
     "    df = pd.read_csv(infile)\n",
     "    \n",
     "    # date\n",
@@ -5731,8 +5731,8 @@
    ],
    "source": [
     "# save\n",
-    "df_train.to_hdf('./data/poliniex_30m.hf',key='train')\n",
-    "df_test.to_hdf('./data/poliniex_30m.hf',key='test')\n",
+    "df_train.to_hdf('./data/poloniex_30m.hf',key='train')\n",
+    "df_test.to_hdf('./data/poloniex_30m.hf',key='test')\n",
     "df_train"
    ]
   },

keras-ddpg.ipynb

@@ -164,7 +164,7 @@
    },
    "outputs": [],
    "source": [
-    "df_train = pd.read_hdf('./data/poliniex_30m.hf',key='train')\n",
+    "df_train = pd.read_hdf('./data/poloniex_30m.hf',key='train')\n",
     "env = PortfolioEnv(\n",
     "    df=df_train,\n",
     "    steps=30, \n",
@@ -176,7 +176,7 @@
     ")\n",
     "env.seed = 0   \n",
     "\n",
-    "df_test = pd.read_hdf('./data/poliniex_30m.hf',key='test')\n",
+    "df_test = pd.read_hdf('./data/poloniex_30m.hf',key='test')\n",
     "env_test = PortfolioEnv(\n",
     "    df=df_test,\n",
     "    steps=30, \n",

readme.md

@@ -8,8 +8,7 @@ The main differences from Jian et. al. 2017 are:
 
 - The first step in a deep learning project should be to make sure the model can overfit, this provides a sanity check. So I am first trying to acheive good results with no trading costs.
 - I have not used portfolio vector memory, which could lead to it incurring large trading costs. But as I have disabled trading costs this shouldn't be a problem.
-- I added some random shifts as data augmentation to prevent overfitting
-- As well as vanilla policy gradient I tried generalized advantage estimation and Deep DPG ([Lillicrap et al. 2015]( http://arxiv.org/pdf/1509.02971v2.pdf))
+- Instead of DPG (Deterinate policy gradient) I tried and DDPG (deep determinate policy gradient ([Lillicrap et al. 2015]( http://arxiv.org/pdf/1509.02971v2.pdf))) and VPG (vanilla policy gradient) with generalized advantage estimation.
 - I tried to replicate the best CNN model from the paper (not the LSTM or RNN models)
 
 Author: wassname
@@ -34,7 +33,7 @@ I have not managed to overfit to the training data or generalise to the test dat
 - enviroments/portfolio.py - contains an openai environment for porfolio trading
 - tensorforce-VPG.ipynb - notebook to try a policy gradient agent
 - keras-ddpg - notebook to try a Deep DPG agent
-- data/polinies_30m.hdf - hdf file with cryptocurrency 30 minutes prices
+- data/poloniex_30m.hdf - hdf file with cryptocurrency 30 minutes prices
 
 # Tests
 

requirements/requirements.txt

@@ -14,3 +14,4 @@ numpy==1.12.1
 matplotlib==2.0.0
 # misc
 tqdm==4.11.2
+seaborn==0.7.1

src/environments/portfolio.py

@@ -17,7 +17,7 @@ class DataSrc(object):
         DataSrc.
 
         df - csv for data frame index of timestamps
-             and multi-index columns levels=[['LTCBTC'],...],['close',...]]
+             and multi-index columns levels=[['LTCBTC'],...],['open','low','high','close']]
         steps - total steps in episode
         scale - scale the data for each episode
         augment - fraction to augment the data by
@@ -167,8 +167,8 @@ class PortfolioEnv(gym.Env):
         An environment for financial portfolio management.
 
         Params:
-            df - data frame index of timestamps
-                 and multi-index columns levels=[['LTCBTC'],...],['close',...]]
+            df - csv for data frame index of timestamps
+                 and multi-index columns levels=[['LTCBTC'],...],['open','low','high','close']]
             steps - steps in episode
             scale - scale data and each episode (except return)
             augment - fraction to randomly shift data by

tensorforce-VPG.ipynb

@@ -183,7 +183,7 @@
     }
    ],
    "source": [
-    "df_train = pd.read_hdf('./data/poliniex_30m.hf',key='train')\n",
+    "df_train = pd.read_hdf('./data/poloniex_30m.hf',key='train')\n",
     "env = EnvWrapper(\n",
     "    df=df_train,\n",
     "    steps=30, \n",
@@ -195,7 +195,7 @@
     ")\n",
     "env.seed = 0   \n",
     "\n",
-    "df_test = pd.read_hdf('./data/poliniex_30m.hf',key='test')\n",
+    "df_test = pd.read_hdf('./data/poloniex_30m.hf',key='test')\n",
     "env_test = EnvWrapper(\n",
     "    df=df_test,\n",
     "    steps=30, \n",

test/test_env.py

@@ -3,7 +3,7 @@ import numpy as np
 from src.environments.portfolio import PortfolioEnv
 
 def test_env_outputs():
-    df = pd.read_hdf('./data/poliniex_30m.hf', key='train')
+    df = pd.read_hdf('./data/poloniex_30m.hf', key='train')
     env = PortfolioEnv(df=df)
 
     action = np.random.random(env.action_space.shape)
@@ -19,7 +19,7 @@ def test_env_outputs():
         assert np.isfinite(v), '%s=%s should be finite' % (k, v)
 
 def test_portfolio_env():
-    df = pd.read_hdf('./data/poliniex_30m.hf', key='train')
+    df = pd.read_hdf('./data/poloniex_30m.hf', key='train')
     asset_names = df.columns.levels[0]
 
     env = PortfolioEnv(df=df)
@@ -41,7 +41,7 @@ def test_portfolio_env():
 
 
 def test_portfolio_env_hold():
-    df = pd.read_hdf('./data/poliniex_30m.hf', key='train')
+    df = pd.read_hdf('./data/poloniex_30m.hf', key='train')
     asset_names = df.columns.levels[0]
 
     np.random.seed(0)
@@ -57,7 +57,7 @@ def test_portfolio_env_hold():
 
 
 def test_scaled():
-    df = pd.read_hdf('./data/poliniex_30m.hf', key='train')
+    df = pd.read_hdf('./data/poloniex_30m.hf', key='train')
 
     np.random.seed(0)
     env1 = PortfolioEnv(df=df, scale=True)