def OLStraining(train,test,t_index,y_true):
df_train = train.copy()
df_test = test.copy()
col = [c for c in df_train.columns if c not in excl]
X = np.array([])
y = np.array([])
X = df_train.loc[df_train.timestamp.isin(t_index), col].copy()
y = df_train.loc[df_train.timestamp.isin(t_index), "y"].copy()
print("--- selected X shape %s" % (X.shape,))
print("--- selected y shape %s" % (y.shape,))
X = X.fillna(X.mean(axis=0))
#
# failed to fill with mean, then fill with zero
#
X = X.fillna(.0)
X = ( X - X.mean(axis=0) ) / X.std(axis=0)
X = np.array(X)
X = sm.add_constant(X)
ols_model = sm.OLS(y,X)
res = ols_model.fit()
print res.summary()
Xt = df_test.loc[:, col].copy()
Xt = Xt.fillna(Xt.mean(axis=0))
Xt = Xt.fillna(.0)
Xt = ( Xt - Xt.mean(axis=0) ) / Xt.std(axis=0)
Xt = np.array(Xt)
Xt = sm.add_constant(Xt)
Xt = sm.add_constant(Xt)
y_pred = res.predict(Xt)
print(len(y_pred),len(y_true))
print r_score(y_true,y_pred)
def testAalysis3(train, test, y_test_true):
df_test = test.copy()
df_train = train.copy()
print("test shape %s" % (df_test.shape, ))
print("y_test_true shape %s" % (y_test_true.shape, ))
uniq_timestamp = df_test["timestamp"].unique()
uniq_id = sorted(df_test["id"].unique())
#print uniq_id
print("-- selected features id: %d" % uniq_timestamp)
train_uniq_id = sorted(df_train["id"].unique())
train_uniq_timestamp = sorted(df_train["timestamp"].unique())
new_test_uniq_id = np.array(
[test_id for test_id in uniq_id if test_id not in train_uniq_id])
print("test timestamp length %d , test id length %d" %
(len(uniq_timestamp), len(uniq_id)))
print("train id length %d" % len(train_uniq_id))
print("length of new test unique id %d" % new_test_uniq_id.shape)
Xtrain = np.array(df_train)
print Xtrain.shape
train_id = Xtrain[:, 0]
scores = {}
for cnt, idx in enumerate(uniq_id):
mask = Xtrain[:, 0] == idx
select_one_id = Xtrain[mask, :].copy()
mask = select_one_id[:, 1] > (uniq_timestamp - 30)
select_one_id = select_one_id[mask, :]
#print select_one_id.shape
data = select_one_id[:, 2:110]
y = select_one_id[:, 110].ravel()
y = np.cumsum(y)
if cnt % 100 == 0:
print("-- trainig counter:%d test_id:%d" % (cnt, idx))
y_pred, final_cost = fitting(data, y)
if not np.isnan(final_cost):
#print idx,r_score(y,y_pred)
scores[idx] = r_score(y, y_pred)
else:
scores[idx] = 0.0
print("-- total counter %d" % cnt)
r_score_value = scores.values()
print np.mean(r_score_value)
def testAalysis4(train, test, y_test_true):
df_test = test.copy()
df_train = train.copy()
print("test shape %s" % (df_test.shape, ))
print("y_test_true shape %s" % (y_test_true.shape, ))
uniq_timestamp = df_test["timestamp"].unique()
uniq_id = sorted(df_test["id"].unique())
#print uniq_id
print("-- selected features timestamp: %d" % uniq_timestamp)
train_uniq_id = sorted(df_train["id"].unique())
train_uniq_timestamp = sorted(df_train["timestamp"].unique())
new_test_uniq_id = np.array(
[test_id for test_id in uniq_id if test_id not in train_uniq_id])
print("test id length %d" % (len(uniq_id)))
print("train id length %d" % len(train_uniq_id))
print("length of new test unique id %d" % new_test_uniq_id.shape)
Xtrain = np.array(df_train)
Xtest = np.array(df_test)
print("* training data shape %s" % (Xtrain.shape, ))
print("* test data shape %s" % (Xtest.shape, ))
train_id = Xtrain[:, 0]
scores = {}
y_test_pred_dict = {}
for cnt, idx in enumerate(uniq_id):
mask = Xtrain[:, 0] == idx
select_one_id = Xtrain[mask, :].copy()
mask = select_one_id[:, 1] > (uniq_timestamp - 30)
select_one_id = select_one_id[mask, :]
#print select_one_id.shape
data = select_one_id[:, 2:110]
y = select_one_id[:, 110].ravel()
y = np.cumsum(y)
if cnt % 100 == 0:
print("-- trainig counter:%d test_id:%d" % (cnt, idx))
#
# data model used from PCA analysis 5 dimentions
#
y_pred, theta, final_cost = fitting(data, y)
#
if not np.isnan(final_cost):
#print idx,r_score(y,y_pred)
scores[idx] = r_score(y, y_pred)
mask = Xtest[:, 0] == idx
select_one_test_id = Xtest[mask, :].copy()
test_data = select_one_test_id[:, 2:]
data_stacked = np.vstack((data[1:, :], test_data))
Xtt = dataPCA(data_stacked)
y_test_diff = np.dot(Xtt, theta)
#y_test_diff = np.diff(y_test)
y_test_prediction = y_test_diff[-1]
y_test_pred_dict[idx] = y_test_prediction
#print(y_test_true[cnt],y_test_prediction)
else:
scores[idx] = 0.0
y_test_pred_dict[idx] = 0.0
print("-- total counter %d" % (cnt + 1))
r_score_value = scores.values()
print np.mean(r_score_value)
return y_test_pred_dict
features = observation.features.copy()
# y_hat = gmodel_test.predict2(observation.features.copy())
target = observation.target
target['y'] = y_hat
timestamp = observation.features["timestamp"][0]
if timestamp % 100 == 0:
print("Timestamp #{}".format(timestamp))
y_true = env.temp_test_y
#y_true = np.exp(y_true)
score_ = r_score(y_true, y_hat)
rewards.append(score_)
print("-- score %.5f" % np.mean(rewards))
print("-- reward %.5f" % reward)
# We perform a "step" by making our prediction and getting back an updated "observation":
observation, reward, done, info = env.step(target)
if done:
print("Public score: {}".format(info["public_score"]))
break
def proc1(log):
env = make()
observation_test = env.reset()
emcv = ElasticNetCV()
#columns = ['technical_30', 'technical_20', 'fundamental_11', 'technical_19']
columns = ['technical_30', 'technical_20', 'fundamental_11']
train_data = observation_test.train.copy()
gmodel_test = glmModel(train_data, columns)
y_hat = gmodel_test.BuildModel()
model_test = fitModel(emcv, train_data, columns)
prediction_test = model_test.predict(observation_test.features.copy())
print "No elasticnet observation :", len(prediction_test)
#score_ = r_score(y_true, y_hat)
#print score_
return 1
"""
train_data = observation_test.train.copy()
features_data = observation_test.features.copy()
feat_colNames = features_data.columns.values.tolist()[2:]
#train_data = observation_test.features.copy
kaggleAnalysis = KaggleDataAnalysisClass(train_data,True)
kaggleAnalysis.corrCheck(feat_colNames)
#emcv = ElasticNetCV(fit_intercept = True)
kaggleAnalysis.modelfit(emcv)
"""
while True:
prediction_test = model_test.predict(observation_test.features.copy())
target_test = observation_test.target
target_test['y'] = prediction_test
"""
features_data = observation_test.features.copy()
prediction_test = kaggleAnalysis.predict(features_data)
target_test = observation_test.target
target_test['y'] = prediction_test
timestamp_ = observation_test.features["timestamp"][0]
log.info("timestamp : %d " % timestamp_)
"""
timestamp_ = observation_test.features["timestamp"][0]
rewards = []
if timestamp_ % 100 == 0:
print(timestamp_)
y_true = env.temp_test_y
score_ = r_score(y_true, prediction_test)
rewards.append(score_)
log.info("score %.5f" % np.mean(rewards))
observation_test, reward_test, done_test, info_test = env.step(
target_test)
#log.info("reward_test : %.5f " % reward_test)
if done_test:
print('Info-test:', info_test['public_score'])
break