前一段时间参加了京东的算法大赛陪跑了一下，试了各种算法，包括lightbgm,xgboost,lstm以及简单的时间序列模型，发现效果最好的竟然是最简单的ARIMA时间序列预测模型，不知道是不是我没有调参还是模型太过简单，因为时间也不是很多，最后那几天真的是很赶，赶着搭建lstm模型（因为听别人说lstm可以达到0.12），但是实际跑出来的结果却不尽人意。最后我们的分数停留在了0.1310，很渣，但是这次比赛我学到了不少东西，尤其是用pandas，numpy处理数据，以及用keras搭建模型。

 

话不多说，具体写一下我的时间序列模型是怎么搭建的吧。本次比赛我们只用到了flow_train.csv的数据，另一个transition 的文件我实在没有想到怎么用，很想看看其他大佬是怎么用的。

1. 数据处理，将原始数据转化为差分，使得数据变得比较稳定。因为时间序列模型只有在数据稳定的情况下才会取得好的效果。
2. 训练策略：将每个district以及每个y值分开训练。

首先，导入我们要用的库：

from statsmodels.tsa.arima_model import ARIMA
import pandas as pd
import numpy as np
import datetime

然后读取flow_train.csv文件：

flow_train = pd.read_csv('/Users/tangshihua/Downloads/jdd/flow_train.csv')

创建保存结果的dataframe：

labels = ['dwell','flow_in','flow_out']star = '20180301'
dates = []
for i in range(1,16):date_format = datetime.datetime.strptime(star,'%Y%m%d')fut_date = date_format + datetime.timedelta(days=i)dates.append(int(datetime.datetime.strftime(fut_date,'%Y%m%d')))test_df = pd.DataFrame({'date_dt':dates})
result = pd.DataFrame(columns=['date_dt', 'city_code', 'district_code', 'dwell', 'flow_in', 'flow_out'])

导入ARIMA模型，fit进train数据：

# train autoregression
model = ARIMA(train,order=(4,1,0))
model_fit = model.fit(maxlag=6, disp=False)

用训练好的模型坐预测：

predictions = model_fit.predict(start=len(train), end=len(train) + 14)
#print(predictions)
#train_set的最后一个值
last_obs = train_set.values[-1]

写进结果那一列：
test_df[y] = predictions

完整的代码在这里：

from statsmodels.tsa.arima_model import ARIMA
import pandas as pd
import numpy as np
import datetime# create a difference transform of the dataset
def difference(dataset):diff = list()for i in range(1, len(dataset)):value = dataset[i] - dataset[i - 1]diff.append(value)return np.array(diff)# Make a prediction give regression coefficients and lag obs
def predict(coef, history):yhat = coef[0]for i in range(1, len(coef)):yhat += coef[i] * history[-i]return yhatflow_train = pd.read_csv('./flow_train.csv')district_set = flow_train.district_code.unique()
# print(district_set)
labels = ['dwell','flow_in','flow_out']star = '20180301'
dates = []
for i in range(1,16):date_format = datetime.datetime.strptime(star,'%Y%m%d')fut_date = date_format + datetime.timedelta(days=i)dates.append(int(datetime.datetime.strftime(fut_date,'%Y%m%d')))test_df = pd.DataFrame({'date_dt':dates})
result = pd.DataFrame(columns=['date_dt', 'city_code', 'district_code', 'dwell', 'flow_in', 'flow_out'])for district in district_set:city = flow_train[flow_train['district_code'] == district].city_code.unique()test_df['city_code'] = city[0]test_df['district_code'] = districtfor y in labels:train_set = flow_train[flow_train['district_code'] == district][[y]]X = difference(train_set.values)size = int(len(X) * 1)train, test = X[0:size], X[size:]# train autoregressionmodel = ARIMA(train,order=(4,1,0))model_fit = model.fit(maxlag=6, disp=False)window = model_fit.k_arcoef = model_fit.params# walk forward over time steps in testhistory = [train[i] for i in range(len(train))]pred = []*15predictions = model_fit.predict(start=len(train), end=len(train) + 14)#print(predictions)last_obs = train_set.values[-1]#print(last_obs)#print(predictions[0])for i in range(len(predictions)):if i == 0 :predictions[i] = last_obs[0] + predictions[i]else:predictions[i] = predictions[i-1] + predictions[i]test_df[y] = predictionsresult = pd.concat([result, test_df[result.columns]])
result['date_dt'] = result['date_dt'].astype(int)
result = result[['date_dt', 'city_code', 'district_code', 'dwell', 'flow_in', 'flow_out']]
result.to_csv('./prediction.csv', index=False, header=None)