毕业设计，在缺失数据的情况下训练arma模型.zip

# -*- coding: utf-8 -*- # author: rw # E-mail: weiyanjie10@gmail.com import numpy as np import copy from sklearn import linear_model from arma import ARMA import random class ArPredicter(object, ): def __init__(self, p, time_series, missing_indexs, stop_error=0.0005, max_iter_time=6): self.p = p self.time_series = time_series self.missing_indexs = missing_indexs self.max_iter_time = max_iter_time self.stop_error = stop_error self.fit() def fit(self): iter_time = 0 while iter_time < self.max_iter_time: clf, error = self.get_clf() if error < self.stop_error: self.clf = clf return iter_time += 1 print iter_time,"th complete" self.clf = clf return def get_clf(self): def get_x_y(time_series): index = 0 while index + self.p < len(time_series): x = time_series[index:index+self.p] y = time_series[index+self.p] index += 1 yield x,y xs_ys = [x_y for x_y in get_x_y(self.time_series)] xs = [x for x,y in xs_ys] ys = [y for x,y in xs_ys] clf = linear_model.LinearRegression() print "training" clf.fit(xs, ys) def impute(index): xs = self.time_series[index-self.p:index] predict_y = clf.predict(xs) error = abs(self.time_series[index] - predict_y) self.time_series[index] = predict_y[0] return error print "imputing" errors = map(impute, self.missing_indexs) return clf, max(errors) @property def ses(self): ses = [] for index, x in enumerate(self.time_series): if index < self.p: continue if index in self.missing_indexs: continue xs = self.time_series[index-self.p:index] y = self.clf.predict(xs) ses.append((y-self.time_series[index])**2) return ses if __name__ == "__main__": a = ARMA([0.3, -0.4, 0.4, -0.5, 0.6], [], 0.1) time_series = [a.generater.next() for i in range(4000)] missing_rate = 0.1 if_missing = [1 if random.random() > missing_rate or time_series.index(t) < 6 else 0 for t in time_series] time_series = [t if flag==1 else 0 for t,flag in zip(time_series,if_missing)] missing_indexs = [index for index, flag in enumerate(if_missing) if flag == 0] p = ArPredicter(5, time_series, missing_indexs) print sum(p.ses)/4000