def test_corr():
"Test stats.corr"
ds = datasets.get_uts()
y = ds.eval("uts.x[:,:3]")
x = ds.eval('Y.x')
n_cases = len(y)
df = n_cases - 2
corr = stats.corr(y, x)
p = stats.rtest_p(corr, df)
for i in range(len(corr)):
r_sp, p_sp = scipy.stats.pearsonr(y[:, i], x)
assert corr[i] == pytest.approx(r_sp)
assert p[i] == pytest.approx(p_sp)
# NaN
with warnings.catch_warnings(): # divide by 0
warnings.simplefilter("ignore")
assert stats.corr(np.arange(10), np.zeros(10)) == 0
# perm
y_perm = np.empty_like(y)
for perm in permute_order(n_cases, 2):
y_perm[perm] = y
stats.corr(y, x, corr, perm)
for i in range(len(corr)):
r_sp, _ = scipy.stats.pearsonr(y_perm[:, i], x)
assert corr[i] == pytest.approx(r_sp)
def test_corr():
"Test stats.corr"
ds = datasets.get_uts()
y = ds.eval("uts.x[:,:3]")
x = ds.eval('Y.x')
n_cases = len(y)
df = n_cases - 2
corr = stats.corr(y, x)
p = stats.rtest_p(corr, df)
for i in xrange(len(corr)):
r_sp, p_sp = scipy.stats.pearsonr(y[:, i], x)
assert_almost_equal(corr[i], r_sp)
assert_almost_equal(p[i], p_sp)
# NaN
r = stats.corr(np.arange(10), np.zeros(10))
eq_(r, 0)
# perm
y_perm = np.empty_like(y)
for perm in permute_order(n_cases, 2):
y_perm[perm] = y
stats.corr(y, x, corr, perm)
for i in xrange(len(corr)):
r_sp, _ = scipy.stats.pearsonr(y_perm[:, i], x)
assert_almost_equal(corr[i], r_sp)
def test_corr():
"Test stats.corr"
ds = datasets.get_uts()
y = ds.eval("uts.x[:,:3]")
x = ds.eval('Y.x')
n_cases = len(y)
df = n_cases - 2
corr = stats.corr(y, x)
p = stats.rtest_p(corr, df)
for i in range(len(corr)):
r_sp, p_sp = scipy.stats.pearsonr(y[:, i], x)
assert_almost_equal(corr[i], r_sp)
assert_almost_equal(p[i], p_sp)
# NaN
with warnings.catch_warnings(): # divide by 0
warnings.simplefilter("ignore")
eq_(stats.corr(np.arange(10), np.zeros(10)), 0)
# perm
y_perm = np.empty_like(y)
for perm in permute_order(n_cases, 2):
y_perm[perm] = y
stats.corr(y, x, corr, perm)
for i in range(len(corr)):
r_sp, _ = scipy.stats.pearsonr(y_perm[:, i], x)
assert_almost_equal(corr[i], r_sp)
def _corr_all_orig(self, pref):
df = []
for dim in self.myexp.dims:
dim_data = load(pref='dis', exp=self.myexp.exp, suffix=dim)[dim]
if dim_data.ndim == 3:
dim_data = np.mean(dim_data, axis=0)
for depth, model_name in self.myexp.models:
self.myexp.set_model(model_name)
dis = self.myexp.dissimilarity()
layer = dis.keys()[-1]
dis = dis[layer]
corr = stats.corr(dis, dim_data, sel='upper')
if self.myexp.bootstrap:
print('bootstrapping stats...')
bf = stats.bootstrap_resample(
dis,
dim_data,
func=stats.corr,
ci=None,
seed=0,
sel='upper',
struct=self.dims[dim].ravel())
for i, b in enumerate(bf):
df.append([dim, depth, model_name, layer, corr, i, b])
else:
df.append([dim, depth, model_name, layer, corr, 0, np.nan])
df = pandas.DataFrame(df,
columns=[
'kind', 'depth', 'models', 'layer',
'correlation', 'iter', 'bootstrap'
])
self.save(df, pref=pref)
return df
def calc(names, data, item, r, d, model=RandomForestRegressor()):
model = make_model(data, names, False, r, d, model)
graph = util.sequence_dotbracket_to_graph(data[item][1], data[item][2])
res = np.array(predict(model, graph))
other = np.array(data[item][0])
res, other = mask(res, other)
value = corr(res, other)[0]
#print '\t',len(data[item][1]),"\t", value
return value
def correlation_all(trees):
from scipy.stats import pearsonr as corr
#We are assuming each data set is normally distributed
tscores_all = np.array([])
rscores_all = np.array([])
for tree in trees:
tscores, rscores = tree.scores()
tscores_all = np.concatenate((tscores_all, tscores))
rscores_all = np.concatenate((rscores_all, rscores))
return corr(tscores_all, rscores_all)
def correlation_all(trees):
from scipy.stats import pearsonr as corr
# We are assuming each data set is normally distributed
tscores_all = np.array([])
rscores_all = np.array([])
for tree in trees:
tscores, rscores = tree.scores()
tscores_all = np.concatenate((tscores_all, tscores))
rscores_all = np.concatenate((rscores_all, rscores))
return corr(tscores_all, rscores_all)
def corr(self):
dis = self.dissimilarity()
df = []
nname = models.NICE_NAMES[self.model_name].lower()
for dim in self.dims:
dim_data = load(pref='dis', exp=self.exp, suffix=dim)
if dim_data is None:
name = self.model_name
self.set_model(dim)
dim_data = self.dissimilarity()
self.set_model(name)
if dim_data is None:
raise Exception('dimension data %s cannot be obtained' %
dim)
dim_data = dim_data[dim]
if dim_data.ndim == 3:
dim_data = np.mean(dim_data, axis=0)
struct = self.dims[dim] if self.exp in ['fonts', 'stefania'
] else None
if self.filter:
dim_data = dim_data[self.sel][:, self.sel]
struct = None
for layer, data in dis.items():
d = data[self.sel][:, self.sel] if self.filter else data
corr = stats.corr(d, dim_data, sel='upper')
if self.bootstrap:
print('bootstrapping stats...')
bf = stats.bootstrap_resample(d,
dim_data,
func=stats.corr,
ci=None,
seed=0,
sel='upper',
struct=struct)
for i, b in enumerate(bf):
df.append([dim, nname, layer, corr, i, b])
else:
df.append([dim, nname, layer, corr, 0, np.nan])
df = pandas.DataFrame(df,
columns=[
'kind', 'models', 'layer', 'correlation',
'iter', 'bootstrap'
])
self.save(df, pref='corr')
if self.task == 'run':
self.plot_single(df, 'corr')
return df
def get_corr(ts1, ts2, period):
'''
ts1 and ts2 are the two traces for computing the correlation
period is how long the trace is in unit of day
'''
# resample the time-series to have equal length
if len(ts1)==len(ts2):
pass
elif len(ts1)>len(ts2):
ts1 = resample(ts1,len(ts2))
else:
ts2 = resample(ts2,len(ts1))
# parameter for filter
fs = float(len(ts1))/(period*24*3600)
lowcut = 1.0/(6*3600) #6hrs
highcut = 1.0/(20*60) #20mins
out_ts1 = butter_bandpass_filter(ts1, lowcut, highcut, fs, order=3)
out_ts2 = butter_bandpass_filter(ts2, lowcut, highcut, fs, order=3)
#return the corr-coef
return corr(out_ts1, out_ts2)[0]
def write_sig_data(round, clim_data, hindcast, ensemble):
from scipy.stats import pearsonr as corr
rho, p = corr(clim_data, hindcast)
#_Write the round and model correlation
line = '%i,%.2f,' % (round, rho)
#_Get whole model hss and rpss, write to file.
hss = HSS(clim_data, hindcast)
rpss = RPSS(clim_data, ensemble)
line = line + '%.2f,%.2f,' % (rpss['all'], hss['all'])
#_Loop through driest/wettest 10/20/30 years and write information
for n in [10, 20, 30]:
hss = HSS(clim_data, hindcast, n, n )
rpss = RPSS(clim_data, ensemble, n, n )
line = line + '%.2f,%.2f,%.2f,%.2f,' % \
(rpss['dry'], rpss['wet'], hss['dry'], hss['wet'])
#_Remove the last comma and return to newline
line = line[:-1] + '\n'
return line
def corr(self):
dis = self.dissimilarity()
df = []
nname = models.NICE_NAMES[self.model_name].lower()
for dim in self.dims:
dim_data = load(pref="dis", exp=self.exp, suffix=dim)
if dim_data is None:
name = self.model_name
self.set_model(dim)
dim_data = self.dissimilarity()
self.set_model(name)
if dim_data is None:
raise Exception("dimension data %s cannot be obtained" % dim)
dim_data = dim_data[dim]
if dim_data.ndim == 3:
dim_data = np.mean(dim_data, axis=0)
struct = self.dims[dim] if self.exp in ["fonts", "stefania"] else None
if self.filter:
dim_data = dim_data[self.sel][:, self.sel]
struct = None
for layer, data in dis.items():
d = data[self.sel][:, self.sel] if self.filter else data
corr = stats.corr(d, dim_data, sel="upper")
if self.bootstrap:
print("bootstrapping stats...")
bf = stats.bootstrap_resample(
d, dim_data, func=stats.corr, ci=None, seed=0, sel="upper", struct=struct
)
for i, b in enumerate(bf):
df.append([dim, nname, layer, corr, i, b])
else:
df.append([dim, nname, layer, corr, 0, np.nan])
df = pandas.DataFrame(df, columns=["kind", "models", "layer", "correlation", "iter", "bootstrap"])
self.save(df, pref="corr")
if self.task == "run":
self.plot_single(df, "corr")
return df
def _corr_all_orig(self, pref):
df = []
for dim in self.myexp.dims:
dim_data = load(pref="dis", exp=self.myexp.exp, suffix=dim)[dim]
if dim_data.ndim == 3:
dim_data = np.mean(dim_data, axis=0)
for depth, model_name in self.myexp.models:
self.myexp.set_model(model_name)
dis = self.myexp.dissimilarity()
layer = dis.keys()[-1]
dis = dis[layer]
corr = stats.corr(dis, dim_data, sel="upper")
if self.myexp.bootstrap:
print("bootstrapping stats...")
bf = stats.bootstrap_resample(
dis, dim_data, func=stats.corr, ci=None, seed=0, sel="upper", struct=self.dims[dim].ravel()
)
for i, b in enumerate(bf):
df.append([dim, depth, model_name, layer, corr, i, b])
else:
df.append([dim, depth, model_name, layer, corr, 0, np.nan])
df = pandas.DataFrame(df, columns=["kind", "depth", "models", "layer", "correlation", "iter", "bootstrap"])
self.save(df, pref=pref)
return df
def test_ml(stock='F',
forecast_out=5,
month=None,
day=None,
year=2019,
plot=False,
volume=False):
# Assume input day is valid trading day
# Want to separate 1 percent of the data to forecast
# Today info
if (month == None or day == None):
today = datetime.datetime.now()
month = today.month
day = today.day
end_date = dt(year, month, day)
trading_days = get_trading_days([2017, 2018, 2019])
end_idx = np.where(end_date == trading_days)[0][0]
end = trading_days[end_idx - forecast_out]
new_start = trading_days[end_idx - forecast_out]
new_end = trading_days[end_idx]
# For prediction
start = datetime.datetime(2016, 4, 1)
df = read_data(stock, start, end)
#df = web.DataReader(stock, 'yahoo', start, end)
#print(df.index)
df = read_data(stock, start, end)
if (df.empty):
#print("SHOULD BE EMPTY")
return [0] * 10, "ERROR"
df = df[df.index <= end]
#print(df.tail(forecast_out))
dfreg = df.loc[:, ['adjusted close', 'volume']]
dfreg['HL_PCT'] = (df['high'] - df['low']) / df['adjusted close'] * 100.0
dfreg['PCT_change'] = (df['adjusted close'] -
df['open']) / df['open'] * 100.0
# For volume testing
if (volume):
dfreg['adjusted close'] = dfreg['volume']
dfreg['EMA'] = get_ema(dfreg, forecast_out)
if (dfreg['EMA'].empty):
return [0] * 10, "ERROR"
dfreg['old close'] = dfreg['adjusted close']
dfreg['adjusted close'] = dfreg['EMA']
# For validation
#print("NEW START: \t{}".format(new_start))
#print("NEW END: \t{}".format(new_end))
#print("VALIDATION START: {} END: {}\n".format(new_start, new_end))
#new_df = web.DataReader(stock, 'yahoo', new_start, new_end)
new_df = read_data(stock, new_start, new_end)
#print("TESTING VALIDATION DATA")
if (new_df.empty):
return [0] * 10, "ERROR"
#print(new_end)
new_df = new_df[new_df.index <= new_end]
#print(new_df)
#exit(1)
new_dfreg = new_df.loc[:, ['adjusted close', 'volume']]
new_dfreg['HL_PCT'] = (new_df['high'] -
new_df['low']) / new_df['adjusted close'] * 100.0
new_dfreg['PCT_change'] = (new_df['adjusted close'] -
new_df['open']) / new_df['open'] * 100.0
# Drop missing value
dfreg.fillna(value=-99999, inplace=True)
new_dfreg.fillna(value=-99999, inplace=True)
# Searating the label here, we want to predict the Adjclose
forecast_col = 'adjusted close'
dfreg['label'] = dfreg[forecast_col].shift(-forecast_out)
X = np.array(dfreg.drop(['label'], 1))
# Scale X for linear regression
X = preprocessing.scale(X)
# Finally want late X and early X for model
X_lately = X[-forecast_out:]
X = X[:-forecast_out]
# Separate label and identify it as y
y = np.array(dfreg['label'])
y = y[:-forecast_out]
# Training and testing sets
X_train = X[:len(X) - forecast_out]
X_test = X[len(X) - forecast_out:]
y_train = y[:len(y) - forecast_out]
y_test = y[len(y) - forecast_out:]
# LinReg
clfreg = LinearRegression(n_jobs=-1)
# QuadReg2
clfpoly2 = make_pipeline(PolynomialFeatures(2), Ridge())
# QuadReg3
clfpoly3 = make_pipeline(PolynomialFeatures(3), Ridge())
# QuadReg4
clfpoly4 = make_pipeline(PolynomialFeatures(4), Ridge())
# QuadReg5
clfpoly5 = make_pipeline(PolynomialFeatures(5), Ridge())
# KNN Regression
clfknn = KNeighborsRegressor(n_neighbors=2)
# Bayesian Ridge
clfbayr = BayesianRidge()
# Neural Network
clfmlp = MLPRegressor(hidden_layer_sizes=(100, 100, 100),
learning_rate='adaptive',
solver='adam',
max_iter=5,
verbose=False)
# Random Forest Regressor
clfrfr = RFR(n_estimators=15)
# Support Vector Regressor
clfsvr = SVR(gamma='auto')
threads = []
models = [
clfreg, clfpoly2, clfpoly3, clfpoly4, clfpoly5, clfknn, clfbayr,
clfrfr, clfsvr
]
fits = [''] * len(models)
for i in range(len(models)):
process = Thread(target=fitting,
args=[models[i], X_train, y_train, fits, i],
name=stock)
process.start()
threads.append(process)
for process in threads:
process.join()
start = time.time()
try:
reg_forecast = fits[0].predict(X_lately)
poly2_forecast = fits[1].predict(X_lately)
poly3_forecast = fits[2].predict(X_lately)
poly4_forecast = fits[3].predict(X_lately)
poly5_forecast = fits[4].predict(X_lately)
try:
knn_forecast = fits[5].predict(X_lately)
except ValueError:
#print("KNN ERROR: {}".format(stock))
#print("F*****g really: {}".format(stock))
#print(X_lately)
#print(X_lately.shape)
knn_forecast = np.zeros(poly5_forecast.shape)
#exit(1)
bayr_forecast = fits[6].predict(X_lately)
rfr_forecast = fits[7].predict(X_lately)
svr_forecast = fits[8].predict(X_lately)
mlp_forecast = fits[6].predict(X_lately)
except AttributeError:
#print("ISSUES WITH {}".format(stock))
return [0] * 10, {}
#print(fits)
#print(threads)
#print(X_train, y_train)
#print(X, y)
#print(stock)
#print(dfreg)
#exit(1)
#mlp_forecast = clfmlp.predict(X_lately)
# Set up dataframe
dfreg['reg_forecast'] = np.nan
dfreg['poly2_forecast'] = np.nan
dfreg['poly3_forecast'] = np.nan
dfreg['poly4_forecast'] = np.nan
dfreg['poly5_forecast'] = np.nan
dfreg['knn_forecast'] = np.nan
dfreg['bayr_forecast'] = np.nan
dfreg['mlp_forecast'] = np.nan
dfreg['rfr_forecast'] = np.nan
dfreg['svr_forecast'] = np.nan
last_date = dfreg.iloc[-1].name
last_unix = last_date
next_unix = last_unix + datetime.timedelta(days=1)
for i in zip(reg_forecast, poly2_forecast, poly3_forecast, poly4_forecast,
poly5_forecast, knn_forecast, bayr_forecast, mlp_forecast,
rfr_forecast, svr_forecast):
next_date = next_unix
next_unix += datetime.timedelta(days=1)
dfreg.loc[next_date] = list(
[np.nan for _ in range(len(dfreg.columns) - 10)] + list(i))
#dfreg['mean_forecast'] = dfreg[['poly2_forecast', 'poly3_forecast']].mean(axis=1)
#print(dfreg.tail(forecast_out+1))
dfreg['mean_forecast'] = dfreg[[
'reg_forecast',
'poly2_forecast',
'poly3_forecast',
'knn_forecast',
'bayr_forecast', # mlp_forecast,
'rfr_forecast',
'svr_forecast'
]].mean(axis=1)
as_list = dfreg.index.tolist()
# I THINK THIS IS FIXED
#print(as_list[-forecast_out-5:])
#for asd in as_list[-forecast_out-1:]:
# print(asd)
#print()
#for asd in new_df.index.tolist():#[:forecast_out]:
# print(asd)
as_list[-forecast_out:] = new_df.index.tolist()[1:]
try:
dfreg.index = as_list
except:
print("DATA MISALIGNMENT FOR: {}".format(stock))
#print(new_df)
#print(dfreg.tail(forecast_out+1))
#exit(1)
return [0] * 10, {}
#for asd in as_list[-forecast_out-5:]:
# print(asd)
dfreg[-forecast_out:].index = new_df.index.tolist()[:forecast_out]
#print(dfreg.tail(forecast_out+1))
#return [None]*10, None
#exit(1)
#
# Trying to do all combinations
#
forecasts = [
'reg_forecast', 'poly2_forecast', 'poly3_forecast', 'poly4_forecast',
'poly5_forecast', 'knn_forecast', 'bayr_forecast', 'rfr_forecast',
'svr_forecast'
]
if (plot):
dfreg['old close'].tail(20).plot(figsize=(20, 12), lw=2)
dfreg['adjusted close'].tail(20).plot(figsize=(20, 12), lw=2)
dfreg['reg_forecast'].tail(20).plot(lw=0.5)
dfreg['poly2_forecast'].tail(20).plot(lw=0.5)
dfreg['poly3_forecast'].tail(20).plot(lw=0.5)
dfreg['poly4_forecast'].tail(20).plot(lw=0.5)
dfreg['poly5_forecast'].tail(20).plot(lw=0.5)
dfreg['knn_forecast'].tail(20).plot(lw=0.5)
dfreg['bayr_forecast'].tail(20).plot(lw=0.5)
dfreg['mean_forecast'].tail(20).plot(c='k')
#dfreg['mlp_forecast'].tail(20).plot()
dfreg['rfr_forecast'].tail(20).plot(lw=0.5)
dfreg['svr_forecast'].tail(20).plot(lw=0.5)
new_dfreg['Actual close'] = new_df['adjusted close']
if (plot):
new_dfreg['Actual close'].tail(20).plot(c='g', lw=2)
fit = np.polyfit([i for i in range(forecast_out)],
dfreg['mean_forecast'].values[-forecast_out:],
deg=1)
#print("CALCULATING CORRELATION BETWEEN METHOD AND ACTUAL")
actual = new_dfreg['Actual close'].tail(forecast_out)
highest_corr = 0
best_comb = ''
num_combs = 0
correlations = []
good_combinations = []
#for j in range(1,9):
# for comb in combinations(forecasts, j):
# num_combs += 1
# comb_dat = dfreg[[*list(comb)]].mean(axis=1).tail(forecast_out)
# new_correlation = corr(comb_dat, actual)[0]
# correlations.append(new_correlation)
# if(new_correlation > 0.4):
# good_combinations.append(comb)
# if(new_correlation > highest_corr):
# highest_corr = new_correlation
# best_comb = comb
for comb in all_combinations:
num_combs += 1
comb_dat = dfreg[[*list(comb)]].mean(axis=1).tail(forecast_out)
new_correlation = corr(comb_dat, actual)[0]
correlations.append(new_correlation)
if (new_correlation > 0.4):
good_combinations.append(comb)
if (new_correlation > highest_corr):
highest_corr = new_correlation
best_comb = comb
reg_dat = dfreg['reg_forecast'].tail(forecast_out)
reg_corr = corr(reg_dat, actual)
#print("Linear Regression: {}".format(reg_corr))
poly2_dat = dfreg['poly2_forecast'].tail(forecast_out)
poly2_corr = corr(poly2_dat, actual)
#print("Poly2: {}".format(poly2_corr))
poly3_dat = dfreg['poly3_forecast'].tail(forecast_out)
poly3_corr = corr(poly3_dat, actual)
#print("Poly3: {}".format(poly3_corr))
poly4_dat = dfreg['poly4_forecast'].tail(forecast_out)
poly4_corr = corr(poly4_dat, actual)
#print("Poly3: {}".format(poly3_corr))
poly5_dat = dfreg['poly5_forecast'].tail(forecast_out)
poly5_corr = corr(poly5_dat, actual)
#print("Poly3: {}".format(poly3_corr))
knn_dat = dfreg['knn_forecast'].tail(forecast_out)
knn_corr = corr(knn_dat, actual)
#print("K Nearest Neighbors: {}".format(knn_corr))
bayr_dat = dfreg['bayr_forecast'].tail(forecast_out)
bayr_corr = corr(bayr_dat, actual)
#print("Bayesian: {}".format(bayr_corr))
rfr_dat = dfreg['rfr_forecast'].tail(forecast_out)
rfr_corr = corr(rfr_dat, actual)
#print("Random Forest: {}".format(rfr_corr))
svr_dat = dfreg['svr_forecast'].tail(forecast_out)
svr_corr = corr(svr_dat, actual)
#print("Support Vector: {}".format(rfr_corr))
mean_dat = dfreg['mean_forecast'].tail(forecast_out)
mean_corr = corr(mean_dat, actual)
if (plot):
plt.legend(loc='best')
plt.xlabel('Date')
plt.ylabel('Price')
plt.title(stock)
plt.savefig("./test_plots/{1}_{2}/{0}_{1}_{2}_{3}".format(
stock, month, day, forecast_out))
plt.close()
return (reg_corr[0], poly2_corr[0], poly3_corr[0], poly4_corr[0], poly5_corr[0],\
knn_corr[0], bayr_corr[0], rfr_corr[0], mean_corr[0], svr_corr[0]), good_combinations
def crossvalpcr(self, xval = True, debug = False):
#Must set phase with bootcorr, and then use crossvalpcr, as it just uses the corr_grid attribute
import numpy as np
from numpy import array
from scipy.stats import pearsonr as corr
from scipy.stats import linregress
from matplotlib import pyplot as plt
from atmos_ocean_data import weightsst
predictand = self.clim_data
if self.corr_grid.mask.sum() >= len(self.sst.lat) * len(self.sst.lon) - 4:
yhat = np.nan
e = np.nan
index = self.clim_data.index
hindcast = pd.Series(data = yhat, index = index)
error = pd.Series(data = e, index = index)
self.correlation = np.nan
self.hindcast = np.nan
self.hindcast_error = np.nan
self.flags['noSST'] = True
return
self.flags['noSST'] = False
sstidx = self.corr_grid.mask == False
n = len(predictand)
yhat = np.zeros(n)
e = np.zeros(n)
idx = np.arange(n)
params = []
std_errs = []
p_vals = []
t_vals = []
if not xval:
rawSSTdata = weightsst(self.sst).data
rawdata = rawSSTdata[:, sstidx]
cvr = np.cov(rawdata.T)
eigval, eigvec = np.linalg.eig(cvr)
eigvalsort = np.argsort(eigval)[::-1]
eigval = eigval[eigvalsort]
eigval = np.real(eigval)
ncomp = 1
eof_1 = eigvec[:,:ncomp] #_fv stands for Feature Vector, in this case EOF-1
eof_1 = np.real(eof_1)
pc_1 = eof_1.T.dot(rawdata.T).squeeze()
self.pc1 = pc_1
return pc_1
for i in idx:
test = idx == i
train = idx != i
rawSSTdata = weightsst(self.sst).data[train]
droppedSSTdata = weightsst(self.sst).data[test]
rawdata = rawSSTdata[:, sstidx]#
dropped_data = droppedSSTdata[:,sstidx].squeeze()
#U, s, V = np.linalg.svd(rawdata)
#pc_1 = V[0,:] #_Rows of V are principal components
#eof_1 = U[:,0].squeeze() #_Columns are EOFS
#EIGs = s**2 #_s is square root of eigenvalues
cvr = np.cov(rawdata.T)
eigval, eigvec = np.linalg.eig(cvr)
eigvalsort = np.argsort(eigval)[::-1]
eigval = eigval[eigvalsort]
eigval = np.real(eigval)
ncomp = 1
eof_1 = eigvec[:,:ncomp] #_fv stands for Feature Vector, in this case EOF-1
eof_1 = np.real(eof_1)
pc_1 = eof_1.T.dot(rawdata.T).squeeze()
slope, intercept, r_value, p_value, std_err = linregress(pc_1, predictand[train])
predictor = dropped_data.dot(eof_1)
yhat[i] = slope * predictor + intercept
e[i] = predictand[i] - yhat[i]
params.append(slope); std_errs.append(std_err); p_vals.append(p_value)
t_vals.append(slope/std_err)
r, p = corr(predictand, yhat)
index = self.clim_data.index
hindcast = pd.Series(data = yhat, index = index)
error = pd.Series(data = e, index = index)
self.hindcast = hindcast
self.hindcast_error = error
self.correlation = round(r, 2)
self.reg_stats = { 'params' : array(params),
'std_errs' : array(std_errs),
't_vals' : array(t_vals),
'p_vals' : array(p_vals)}
return
def crossvalpcr(self, fig = None, ax = None, phase = 'allyears', onlySST = False, debug = False):
#Must set phase with bootcorr, and then use crossvalpcr, as it just uses the corr_grid attribute
import numpy as np
from scipy.stats import pearsonr as corr
from scipy.stats import linregress
from matplotlib import pyplot as plt
"""
if fig == None:
fig = plt.figure()
ax = fig.add_subplot(111)
"""
#Set up predictand from climate data
predictand = self.clim_data[phase]
#predictand = (predictand - predictand.mean())/predictand.std()
#self.predictand[phase] = predictand
#Get an index of every significantly correlated gridpoint for the predictor fields
if self.corr_grid['sst'][phase].mask.sum() >= 16019:
print ('No sig SST or SLP')
yhat = np.zeros(len(self.clim_data[phase]))
e = np.zeros(len(self.clim_data[phase]))
index = self.clim_data[phase].index
hindcast = pd.Series(data = yhat, index = index)
error = pd.Series(data = e, index = index)
self.hindcast[phase] = hindcast
self.hindcast_error[phase] = error
self.nosigSST = True
return
else:
self.nosigSST = False
sstidx = self.corr_grid['sst'][phase].mask == False
# if self.corr_grid['slp'][phase].mask.sum() == 2664:
# onlySST = True
if not onlySST:
slpidx = self.corr_grid['slp'][phase].mask == False
#Set up some empty variables
n = len(predictand)
yhat = np.zeros(n)
e = np.zeros(n)
idx = np.arange(n)
for i in idx:
test = idx == i
train = idx != i
rawSSTdata = self.sst[phase][train]
rawSLPdata = self.slp[phase][train]
droppedSSTdata = self.sst[phase][test]
droppedSLPdata = self.slp[phase][test]
if onlySST:
rawdata = rawSSTdata[:, sstidx]#.T
dropped_data = droppedSSTdata[:,sstidx].squeeze()
else:
rawdata = np.concatenate((rawSSTdata[:, sstidx], rawSLPdata[:, slpidx]), axis = 1) #.T
dropped_data = np.concatenate((droppedSSTdata[:,sstidx], droppedSLPdata[:,slpidx]), axis = 1)#.T.squeeze()
#U, s, V = np.linalg.svd(rawdata)
#pc_1 = V[0,:] #_Rows of V are principal components
#eof_1 = U[:,0].squeeze() #_Columns are EOFS
#EIGs = s**2 #_s is square root of eigenvalues
cvr = np.cov(rawdata.T)
eigval, eigvec = np.linalg.eig(cvr)
eigvalsort = np.argsort(eigval)[::-1]
eigval = eigval[eigvalsort]
eigval = np.real(eigval)
ncomp = 1
eof_1 = eigvec[:,:ncomp] #_fv stands for Feature Vector, in this case EOF-1
eof_1 = np.real(eof_1)
pc_1 = eof_1.T.dot(rawdata.T).squeeze()
slope, intercept, r_value, p_value, std_err = linregress(pc_1, predictand[train])
predictor = dropped_data.dot(eof_1)
yhat[i] = slope * predictor + intercept
e[i] = predictand[i] - yhat[i]
c = corr(predictand, yhat)
"""
ax.scatter(predictand, yhat)
ax.set_title('%s, r = %f' % (phase, round(c[0],2)))
ax.axis([0,15,0,15])
"""
index = self.clim_data[phase].index
hindcast = pd.Series(data = yhat, index = index)
error = pd.Series(data = e, index = index)
self.hindcast[phase] = hindcast
self.hindcast_error[phase] = error
return fig, ax
def bootcorr(self, n = 100, fig = None, ax = None, field = 'sst', \ phase = 'allyears', corrconf = 0.9, bootconf = 0.9, cbloc = 'bottom',\ quick = False, debug = False, monte = False): from numpy import meshgrid, zeros, ma, isnan, linspace import time from random import sample if field == 'sst': fieldData = self.sst[phase] if field == 'slp': fieldData = self.slp[phase] clim_data = self.clim_data[phase] corrlevel = 1 - corrconf corr_grid = vcorr(X = fieldData, y = clim_data) n_yrs = len(clim_data) p_value = sig_test(corr_grid, n_yrs) #Mask insignificant gridpoints corr_grid = ma.masked_array(corr_grid, ~(p_value < corrlevel)) #Mask land corr_grid = ma.masked_array(corr_grid, isnan(corr_grid)) #Mask northern/southern ocean corr_grid.mask[self.lat[field] > 60] = True corr_grid.mask[self.lat[field] < -30] = True ###SAVE THE MASK TO FILTER THE BOOTSTRAP mask = corr_grid.mask if quick: self.corr_grid[field][phase] = corr_grid return ###SET UP INDICES FOR FIELD DATA### ###INITIALIZE A NEW CORR GRID#### nlat = fieldData.shape[1] nlon = fieldData.shape[2] count = np.zeros((nlat,nlon)) ntim = n dat = clim_data mask = corr_grid.mask if debug: print 'Starting %s' % phase for boot in xrange(ntim): if debug: print 'starting round %i' % boot ###SHUFFLE THE YEARS AND CREATE THE BOOT DATA### idx = np.random.randint(0, len(dat) - 1, len(dat)) bootdata = np.zeros((len(idx), nlat, nlon)) bootdata[:] = fieldData[idx] bootvar = np.zeros((len(idx))) bootvar = dat[idx] corr_grid_boot = vcorr(X = bootdata, y = bootvar) n_yrs = len(bootvar) p_value = sig_test(corr_grid_boot, n_yrs) count[p_value <= corrlevel] += 1 if debug: print 'Count max is %i' % count.max() # for lon, lat in zip(xx[~mask], yy[~mask]): # c, p = corr(bootdata[:,lat,lon], bootvar) # if p <= corrlevel: # count[lat,lon] += 1 ###GET THE ACTUAL CORRELATION AGAIN #_Mask insignificant values ###CREATE MASKED ARRAY USING THE COUNT AND BOOTCONF ATTRIBUTES corr_grid = np.ma.masked_array(corr_grid, count < bootconf * ntim) self.corr_grid[field][phase] = corr_grid if monte: n_phase = len(clim_data) n_total = len(self.clim_data['allyears']) count = np.zeros((nlat, nlon)) field = self.sst[phase] for t in xrange(100): #print 'Starting monte round %i' % t idx = sample(xrange(n_total), n_phase) var = self.clim_data['allyears'][idx] for lon, lat in zip(xx[~mask], yy[~mask]): r, p = corr(var, field[:,lat,lon]) if p <= (1 - corrconf): #print 'Entering while loop for grid %i, %i' % (i, j) x = 0 c2 = 0 while x < 100: #print 'Monte round %i' % x idx2 = np.random.randint(0, n_phase-1, n_phase) x += 1 r, p = corr(var[idx2], field[idx2,lat,lon]) if p <= (1 - corrconf): c2 += 1 if c2 >= 80: count[lat,lon] += 1 print 'Count max is %.0f' % (count.max()) self.monte_count[phase] = count.max() self.monte_grid[phase] = np.ma.masked_array(data = count, \ mask = self.corr_grid['sst'][phase].mask) return
def crossvalpcr(self, xval=True, debug=False):
#Must set phase with bootcorr, and then use crossvalpcr, as it just uses the corr_grid attribute
import numpy as np
from numpy import array
from scipy.stats import pearsonr as corr
from scipy.stats import linregress
from matplotlib import pyplot as plt
from utils import weightsst
predictand = self.clim_data
if self.corr_grid.mask.sum(
) >= len(self.sst.lat) * len(self.sst.lon) - 4:
yhat = np.nan
e = np.nan
#index = self.clim_data.index
index = self.mei
hindcast = pd.Series(data=yhat, index=index)
error = pd.Series(data=e, index=index)
self.correlation = np.nan
self.hindcast = np.nan
self.hindcast_error = np.nan
self.flags['noSST'] = True
return
self.flags['noSST'] = False
sstidx = self.corr_grid.mask == False
n = len(predictand)
yhat = np.zeros(n)
e = np.zeros(n)
idx = np.arange(n)
params = []
std_errs = []
p_vals = []
t_vals = []
if not xval:
rawSSTdata = weightsst(self.sst).data
rawdata = rawSSTdata[:, sstidx]
cvr = np.cov(rawdata.T)
eigval, eigvec = np.linalg.eig(cvr)
eigvalsort = np.argsort(eigval)[::-1]
eigval = eigval[eigvalsort]
eigval = np.real(eigval)
ncomp = 1
eof_1 = eigvec[:, :
ncomp] #_fv stands for Feature Vector, in this case EOF-1
eof_1 = np.real(eof_1)
pc_1 = eof_1.T.dot(rawdata.T).squeeze()
slope, intercept, r, p, err = linregress(pc_1, predictand)
yhat = slope * pc_1 + intercept
self.pc1 = pc_1
self.correlation = r
self.hindcast = yhat
return
for i in idx:
test = idx == i
train = idx != i
rawSSTdata = weightsst(self.sst).data[train]
droppedSSTdata = weightsst(self.sst).data[test]
rawdata = rawSSTdata[:, sstidx] #
dropped_data = droppedSSTdata[:, sstidx].squeeze()
#U, s, V = np.linalg.svd(rawdata)
#pc_1 = V[0,:] #_Rows of V are principal components
#eof_1 = U[:,0].squeeze() #_Columns are EOFS
#EIGs = s**2 #_s is square root of eigenvalues
cvr = np.cov(rawdata.T)
#print cvr.shape
eigval, eigvec = np.linalg.eig(cvr)
eigvalsort = np.argsort(eigval)[::-1]
eigval = eigval[eigvalsort]
eigval = np.real(eigval)
ncomp = 1
eof_1 = eigvec[:, :
ncomp] #_fv stands for Feature Vector, in this case EOF-1
eof_1 = np.real(eof_1)
pc_1 = eof_1.T.dot(rawdata.T).squeeze()
slope, intercept, r_value, p_value, std_err = linregress(
pc_1, predictand[train])
predictor = dropped_data.dot(eof_1)
yhat[i] = slope * predictor + intercept
e[i] = predictand[i] - yhat[i]
params.append(slope)
std_errs.append(std_err)
p_vals.append(p_value)
t_vals.append(slope / std_err)
r, p = corr(predictand, yhat)
hindcast = yhat
error = e
self.hindcast = hindcast
self.hindcast_error = error
self.correlation = round(r, 2)
self.reg_stats = {
'params': array(params),
'std_errs': array(std_errs),
't_vals': array(t_vals),
'p_vals': array(p_vals)
}
return
def ksi_func(pp, tt):
ppx, ppy = pp.T
rat_x = corr(ppx, tt.flatten())[0] ** 2
rat_y = corr(ppy, tt.flatten())[0] ** 2
return 1./ rat_x + 1./ rat_y
plt.plot(range(1, ph + 1), tmp[::2], c='red', label='ask')
plt.plot(range(1, ph + 1), tmp[1::2], c='green', label='bid')
plt.yscale('log')
plt.xlabel('$h$')
plt.ylabel('Jarque-Bera statistic for $h$-step log returns')
for type in ['pdf', 'png']:
plt.savefig(f'{path_save}/log_returns_normality.{type}',
bbox_inches='tight',
dpi=300)
plt.show()
# %% Correlation
tmp = res
cor_ask = np.array(
[corr(res[:, i], res[:, 0])[0] for i in range(0, res.shape[1], 2)])
cor_bid = np.array(
[corr(res[:, i], res[:, 1])[0] for i in range(1, res.shape[1], 2)])
cor_ask_bid = np.array(
[corr(res[:, i], res[:, 1])[0] for i in range(0, res.shape[1], 2)])
cor_bid_ask = np.array(
[corr(res[:, i], res[:, 0])[0] for i in range(1, res.shape[1], 2)])
plt.plot(range(0, ph), cor_ask, c='red', label='ask')
plt.plot(range(0, ph), cor_bid, c='green', label='bid')
plt.plot(range(0, ph), cor_ask_bid, c='red', label='ask2', linestyle='dashed')
plt.plot(range(0, ph),
cor_bid_ask,
c='green',
def crossvalpcr(self, xval = True, debug = False):
#Must set phase with bootcorr, and then use crossvalpcr, as it just uses the corr_grid attribute
import numpy as np
import statsmodels.api as sm
from numpy import array, ndarray, hstack, zeros, vstack
from scipy.stats import pearsonr as corr
from scipy.stats import linregress
from matplotlib import pyplot as plt
from atmos_ocean_data import weightsst
predictand = self.clim_data.copy()
pcs = self.pcs.copy()
n = len(predictand)
yhat = np.zeros(n)
e = np.zeros(n)
params = []
std_errs = []
p_vals = []
t_vals = []
rnd = 0
ncomps, nt = pcs.shape
selection_index = range(ncomps)
xval_idx = np.arange(nt)
best_score = 0.01
scores = [0]
overall_index = []
final_pcs = zeros((nt))
while best_score >= scores[rnd]:
score = zeros((len(selection_index)))
# print 'Round %i' % rnd
for ind, index in enumerate(selection_index):
# print 'Checking pc-%i: ' % index
# print overall_index + [ind]
data = pcs[overall_index + [ind]]
# if rnd ==3: import pdb; pdb.set_trace()
for i in xval_idx:
test = xval_idx == i
train = xval_idx != i
rawdata = data.T[train]
dropped_data = data.T[test]
X = sm.add_constant(data.T[train])
y = predictand[train]
olsmod = sm.OLS(y, X)
olsres = olsmod.fit()
intercept = array([[0]])
yhat[i] = olsres.predict(hstack((intercept, dropped_data)))
e[i] = predictand[i] - yhat[i]
# params.append(slope); std_errs.append(std_err); p_vals.append(p_value)
# t_vals.append(slope/std_err)
r, p = corr(predictand, yhat)
score[ind] = abs(r)
# print 'score is %.2f' % abs(r)
if max(score) > best_score:
best_score = max(score)
best_loc = np.where(score == max(score))[0][0]
best_pc = selection_index[best_loc]
overall_index.append(best_pc)
# print overall_index
# print pcs.shape
selection_index = np.delete(selection_index, best_loc)
# print selection_index
final_pcs = vstack((final_pcs, pcs[best_loc]))
scores.append(best_score)
else:
break
rnd += 1
self.overall_index = overall_index
### NOW REBUILD BEST MODEL ###
data = final_pcs[1:]
#print 'Overall index is ', overall_index
for i in xval_idx:
test = xval_idx == i
train = xval_idx != i
rawdata = data.T[train]
dropped_data = data.T[test]
X = sm.add_constant(data.T[train])
y = predictand[train]
olsmod = sm.OLS(y, X)
olsres = olsmod.fit()
intercept = array([[0]])
yhat[i] = olsres.predict(hstack((intercept, dropped_data)))
e[i] = predictand[i] - yhat[i]
# params.append(slope); std_errs.append(std_err); p_vals.append(p_value)
# t_vals.append(slope/std_err)
###Okay, best first predictor has been obtained... Now need to keep adding
r, p = corr(predictand, yhat)
index = self.clim_data.index
hindcast = pd.Series(data = yhat, index = index)
error = pd.Series(data = e, index = index)
self.hindcast = hindcast
self.hindcast_error = error
self.correlation, _ = corr(self.clim_data, self.hindcast)
# self.reg_stats = { 'params' : array(params),
# 'std_errs' : array(std_errs),
# 't_vals' : array(t_vals),
# 'p_vals' : array(p_vals)}
return
def main():
#To plot yost v theoretical
# arrPitch = []
# arrRoll = []
# arrAx = []
# arrAy = []
# imu = int(input("Enter 0 for YOST IMU, 1 for MPU6050:\n"))
# fileName = input("Enter the acceleration reading file path and name:\n")
experiments = [
"../Data/YOST_stewart_0degPitch_10sPeriod_test_1.txt",
"../Data/MPU6050_stewart_0degPitch_10sPeriod_test_1.txt",
"../Data/YOST_stewart_0degPitch_20sPeriod_test_2.txt",
"../Data/MPU6050_stewart_0degPitch_20Period_test_2.txt",
"../Data/YOST_stewart_20degPitch_20sPeriod_test_3.txt",
"../Data/MPU6050_stewart_20degPitch_20Period_test_3.txt"
]
plot = True
displacements = []
sigWaveHeights = []
for i in range(0, 6):
arrAz = []
totalTime = 0
imu = i % 2 #YOST => 0, MPU6050 => 1
with open(experiments[i]) as f:
#Valid files
#YOST_stewart_0degPitch_10sPeriod_test_1.txt
#YOST_stewart_0degPitch_20sPeriod_test_2.txt
#YOST_stewart_20degPitch_20sPeriod_test_3.txt
#MPU6050_stewart_0degPitch_10sPeriod_test_1.txt
#MPU6050_stewart_0degPitch_20Period_test_2.txt
#MPU6050_stewart_20degPitch_20Period_test_3.txt
#If YOST IMU (imu = 0)
#Data format: "%int(Month)/%int(Day)/%int(Year),%int(Hours):%int(Minutes):%float(Seconds),
# %float(OrientPitch),%float(OrientYaw),%float(OrientRoll),
# %float(CorrectedGyroX),%float(CorrectedGyroY),%float(CorrectedGyroZ),
# %float(CorrectedAccelX),%float(CorrectedAccelY),%float(CorrectedAccelZ),
# %float(CorrectedMagX),%float(CorrectedMagY),%float(CorrectedMagZ)"
if (imu == 0):
f.readline() # Read in first line - this is the Foramt
#Get values from file
startTime = 0
endTime = 0
for line in f:
row = line.split(',')
#Get start time
if (startTime == 0):
startTime = row[0].split(' ')[1]
#Get end time
endTime = row[0].split(' ')[1]
#Select relevent accleration data - comment out if plotting yost v theoretical
row = row[7:10]
#Set upper bound of 0.5g Az
if (float(row[1]) > 0.5 * g):
row[1] = str(0.5 * -g)
arrAz.append(
float(row[1]) *
-g) #comment out if plotting yost v theoretical
#This is also used to compare yost with the true signal
# arrAz.append(float(row[-5])*-g )
# arrAx.append(float(row[-6])*-g )
# arrAy.append(float(row[-4])*-g )
# arrPitch.append(float(row[1]))
# arrRoll.append( float(row[3]) )
#Calculate the sampling frequency
startTime = startTime.split(':')
endTime = endTime.split(':')
totalTime = []
totalTime.append(float(endTime[0]) - float(startTime[0]))
totalTime.append(float(endTime[1]) - float(startTime[1]))
totalTime.append(float(endTime[2]) - float(startTime[2]))
totalTime = totalTime[0] * 60 * 60 + totalTime[
1] * 60 + totalTime[2]
#Else MPU6050 (imu = 1)
#Data format: "int(timeSinceStart ms), float(accelAx mg), float(accelAy mg), float(accelAz g)"
else:
startTime = -1
endTime = 0
for line in f:
#Format is: int ms, float ax, float ay, float az
row = line.split(',')
if (startTime == -1):
startTime = float(row[0]) * 10**-3
endTime = float(row[0]) * 10**-3
#Set upper bound of 0.5g Az
if (float(row[3]) > 0.5 * g):
row[1] = str(0.5 * -g)
#arrAx.append(float(row[1])*-g/1000 )
#arrAy.append(float(row[2])*-g/1000 )
arrAz.append(float(row[3]) * -g)
totalTime = endTime - startTime
fs = len(arrAz) / (totalTime) #Sampling frequency
fs = round(fs) #Account for errors
##Debuging and graphing
#print("Sampling rate = " + str(fs))
#trueVerticalAcceleration(arrAx, arrAy, arrAz, arrPitch,arrRoll, fs)
##EndDebug
#Condition signal:
azFiltered = cond.condition(arrAz, fs, plot)
#Calculate Wave height time series
eta, times = heightTimeSeries(azFiltered, fs, plot, plot, plot)
#Resample to allow for comparison between the imus (has to have same amount of samples)
eta180, times = sig.resample(eta, 180, t=times)
if (plot):
plt.plot(times, eta180, label="Reasmpled heights")
plt.legend(loc='lower right')
plt.show()
displacements.append(eta180)
ht = significantWaveHeight(eta)
hs = spectralSignificantWaveHeight(eta, fs)
sigWaveHeights.append((ht, hs))
# print(displacements)
h = 0.045
c = 0.155
f = 0.1
t = np.arange(0, 90, 0.5)
s = h * np.sin(2 * np.pi * f * t)
for j in range(0, 6):
if (j % 2 == 0):
print("YOST Significant Wave Height (Ht, Hs) for test " +
str(round(j * 2 / 5)) + ": Ht=" +
'{:6f}'.format(sigWaveHeights[j][0] * 1000) + "mm Hs=" +
'{:6f}'.format(sigWaveHeights[j][1] * 1000))
else:
print("MPU6050 Significant Wave Height(Ht, Hs) for test " +
str(round(j * 2 / 5)) + ": Ht=" +
'{:6f}'.format(sigWaveHeights[j][0] * 1000) + "mm Hs=" +
'{:6f}'.format(sigWaveHeights[j][1] * 1000))
print("Theoretical Significant Wave Height: " +
'{:6f}'.format(significantWaveHeight(s) * 1000) + "mm")
for k in range(0, 6, 2):
print("Pearson coerrelation coefficient between IMUs for test " +
str(int(k / 2)) + " is: " + '{:6f}'.format(
abs(corr(displacements[k], displacements[k + 1])[0])))
if(self.dimensionality(n+1)>=dimensionality):
return(n+1)
if __name__ == "__main__":
from scipy.stats import pearsonr as corr
import matplotlib.pyplot as plt
# Test dataset of random value
m, n = 128, 10
weightMatrix = np.random.randn(m, n)
pcaObj = PCA(matrix=weightMatrix)
# Check that we get back the same matrix we put in when we set the number
# of selected components equal to a number of possible values (up to n)
for i in [n,5,1]:
filteredMatrix = pcaObj.filterMatrix(n=i)
pcaObjFiltered = PCA(matrix=filteredMatrix)
isMatrixSame = np.allclose(weightMatrix,filteredMatrix)
r = corr(weightMatrix.flatten(),filteredMatrix.flatten())
print("\nIs PCA Filtered Matrix same (n={})?: {}".format(n,isMatrixSame))
print("Original matrix: {}".format( weightMatrix.flatten()))
print("Filtered matrix: {}".format(filteredMatrix.flatten()))
print("Pearson's Correlation: r = {:4.3f}".format(r[0]))
print('Target Dimensionality at 50%: {}'.format(pcaObj.computeTargetPCs(0.5)))
(fig,ax)=plt.subplots(nrows=2,ncols=1)
ax[0].imshow(weightMatrix.T)
ax[0].set_title('Original dimensionality: {:4.2f}'.format(pcaObj.dimensionality()))
ax[1].imshow(filteredMatrix.T)
ax[1].set_title('Filtered dimensionality: {:4.2f}'.format(pcaObjFiltered.dimensionality()))
