灰色预测实现见：https://www.jianshu.com/p/a35ba96d852b
from pandas import Series
from pandas import DataFrame
import pandas as pd
import matplotlib.pyplot as plt
class Gray_model:
def __init__(self):
self.a_hat = None
self.x0 = None
def fit(self,
series=pd.Series(index=[1996, 1997, 1998, 1999], data=[1, 2, 3, 4])):
"""
Series is a pd.Series with index as its date.
:param series: pd.Series
:return: None
"""
self.a_hat = self._identification_algorithm(series.values)
self.x0 = series.values[0]
def predict(self, interval):
result = []
for i in range(interval):
result.append(self.__compute(i))
result = self.__return(result)
return result
def _identification_algorithm(self, series):
B = np.array([[1] * 2] * (len(series) - 1))
series_sum = np.cumsum(series)
for i in range(len(series) - 1):
B[i][0] = (series_sum[i] + series_sum[i + 1]) * (-1.0) / 2
Y = np.transpose(series[1:])
BT = np.transpose(B)
a = np.linalg.inv(np.dot(BT, B))
a = np.dot(a, BT)
a = np.dot(a, Y)
a = np.transpose(a)
return a
def score(self, series_true, series_pred, index):
error = np.ones(len(series_true))
relativeError = np.ones(len(series_true))
for i in range(len(series_true)):
error[i] = series_true[i] - series_pred[i]
relativeError[i] = error[i] / series_pred[i] * 100
score_record = {'GM': np.cumsum(series_pred),
'1—AGO': np.cumsum(series_true),
'Returnvalue': series_pred,
'Real_value': series_true,
'Error': error,
'RelativeError(%)': (relativeError)
}
scores = DataFrame(score_record, index=index)
return scores
def __compute(self, k):
return (self.x0 - self.a_hat[1] / self.a_hat[0]) * np.exp(-1 * self.a_hat[0] * k) + self.a_hat[1] / self.a_hat[
0]
def __return(self, series):
tmp = np.ones(len(series))
for i in range(len(series)):
if i == 0:
tmp[i] = series[i]
else:
tmp[i] = series[i] - series[i - 1]
return tmp
def evaluate(self, series_true, series_pred):
scores = self.score(series_true, series_pred, np.arange(len(series_true)))
error_square = np.dot(scores, np.transpose(scores))
error_avg = np.mean(error_square)
S = 0 # X0的关联度
for i in range(1, len(series_true) - 1, 1):
S += series_true[i] - series_true[0] + (series_pred[-1] - series_pred[0]) / 2
S = np.abs(S)
SK = 0 # XK的关联度
for i in range(1, len(series_true) - 1, 1):
SK += series_pred[i] - series_pred[0] + (series_pred[-1] - series_pred[0]) / 2
SK = np.abs(SK)
S_Sub = 0 # |S-SK|b
for i in range(1, len(series_true) - 1, 1):
S_Sub += series_true[i] - series_true[0] - (series_pred[i] - series_pred[0]) + ((series_true[-1] -
series_true[0]) - (
series_pred[i] -
series_pred[0])) / 2
S_Sub = np.abs(S_Sub)
T = (1 + S + SK) / (1 + S + SK + S_Sub)
level = 0
if T >= 0.9:
level = 1
# print ('精度为一级')
elif T >= 0.8:
level = 2
# print ('精度为二级')
elif T >= 0.7:
level = 3
# print ('精度为三级')
elif T >= 0.6:
level = 4
# print ('精度为四级')
return 1 - T, level
def plot(self, series_true, series_pred, index):
df = pd.DataFrame(index=index)
df['Real'] = series_true
df['Forcast'] = series_pred
plt.figure()
df.plot(figsize=(7, 5))
plt.xlabel('year')
plt.show()