def _compute_ratio(
self, index: Union[pd.DatetimeIndex,
pd.PeriodIndex]) -> np.ndarray:
if isinstance(index, pd.PeriodIndex):
index = index.to_timestamp()
delta = index.to_perioddelta(self._freq)
pi = index.to_period(self._freq)
gap = (pi + 1).to_timestamp() - pi.to_timestamp()
return to_numpy(delta) / to_numpy(gap)
def in_sample(self) -> pd.DataFrame:
if self._cached_in_sample is not None:
return self._cached_in_sample
index = self._index
if not self._deterministic_terms:
return pd.DataFrame(np.empty((index.shape[0], 0)), index=index)
raw_terms = []
for term in self._deterministic_terms:
raw_terms.append(term.in_sample(index))
raw_terms = self._adjust_dummies(raw_terms)
terms: pd.DataFrame = pd.concat(raw_terms, axis=1)
terms = self._remove_zeros_ones(terms)
if self._drop:
terms_arr = to_numpy(terms)
res = qr(terms_arr, mode="r", pivoting=True)
r = res[0]
p = res[-1]
abs_diag = np.abs(np.diag(r))
tol = abs_diag[0] * terms_arr.shape[1] * np.finfo(float).eps
rank = int(np.sum(abs_diag > tol))
rpx = r.T @ terms_arr
keep = [0]
last_rank = 1
# Find the left-most columns that produce full rank
for i in range(1, terms_arr.shape[1]):
curr_rank = np.linalg.matrix_rank(rpx[:i + 1, :i + 1])
if curr_rank > last_rank:
keep.append(i)
last_rank = curr_rank
if curr_rank == rank:
break
if len(keep) == rank:
terms = terms.iloc[:, keep]
else:
terms = terms.iloc[:, np.sort(p[:rank])]
self._retain_cols = terms.columns
self._cached_in_sample = terms
return terms