+- Python Forum (https://python-forum.io)
+-- Forum: Python Coding (https://python-forum.io/forum-7.html)
+--- Forum: Data Science (https://python-forum.io/forum-44.html)
+--- Thread: Hurst Exponent in Rolling Basis (/thread-31829.html)
Hurst Exponent in Rolling Basis - illmattic - Jan-05-2021
Hello,
I have come across code (https://github.com/Mottl/hurst/blob/master/hurst/__init__.py) that calculates the Hurst exponent from a time series. I have simplified it for my purpose:
def custom_hurst(series):
series = series.tail(365)
max_window = len(series)
min_window = 15
ndarray_likes = [np.ndarray]
if "pandas.core.series" in sys.modules.keys():
ndarray_likes.append(pd.core.series.Series)
# convert series to numpy array if series is not numpy array or pandas Series
if type(series) not in ndarray_likes:
series = np.array(series)
if "pandas.core.series" in sys.modules.keys() and type(series) == pd.core.series.Series:
if series.isnull().values.any():
raise ValueError("Series contains NaNs")
series = series.values # convert pandas Series to numpy array
elif np.isnan(np.min(series)):
raise ValueError("Series contains NaNs")
def to_inc(x):
incs = x[1:] - x[:-1]
return incs
def to_pct(x):
pcts = x[1:] / x[:-1] - 1.
return pcts
def RS_func(series):
incs = to_pct(series)
mean_inc = np.sum(incs) / len(incs)
deviations = incs - mean_inc
Z = np.cumsum(deviations)
R = max(Z) - min(Z)
S = np.std(incs, ddof=1)
return R / S
err = np.geterr()
np.seterr(all='raise')
max_window = max_window or len(series)-1
window_sizes = [15,30,45,90,182,365]
RS = []
for w in window_sizes:
rs = []
for start in range(0, len(series), w):
if (start+w)>len(series):
break
_ = RS_func(series[start:start+w])
if _ != 0:
rs.append(_)
RS.append(np.mean(rs))
A = np.vstack([np.log10(window_sizes), np.ones(len(RS))]).T
H, c = np.linalg.lstsq(A, np.log10(RS), rcond=-1)[0]
np.seterr(**err)
c = 10**c
return HNot being proficient in python, I am struggling to workout a way to have this code applied on a rolling basis to calculate a Hurst value for each previous 365 values in the time series instead of just one Hurst value for the last 365 values. Any help would be appreciated.
Thanks
Matt
RE: Hurst Exponent in Rolling Basis - illmattic - Jan-06-2021
I've tried using the function like this:
df.apply(lambda row: custom_hurst(df), axis=1)
Output:1983-03-30 0.682463
1983-03-31 0.682463
1983-04-04 0.682463
1983-04-05 0.682463
1983-04-06 0.682463
2020-12-29 0.682463
2020-12-30 0.682463
2020-12-31 0.682463
2021-01-04 0.682463
2021-01-05 0.682463