python - 仅选择 Pandas 中每连续分钟都有数据的日期范围

Question

我正在尝试处理 pandas 中的一些数据，这些数据在 CSV 中如下所示:

2014.01.02,08:56,1.37549,1.37552,1.37549,1.37552,3
2014.01.02,09:00,1.37562,1.37562,1.37545,1.37545,21
2014.01.02,09:01,1.37545,1.37550,1.37542,1.37546,18
2014.01.02,09:02,1.37546,1.37550,1.37546,1.37546,15
2014.01.02,09:03,1.37546,1.37563,1.37546,1.37559,39
2014.01.02,09:04,1.37559,1.37562,1.37555,1.37561,37
2014.01.02,09:05,1.37561,1.37564,1.37558,1.37561,35
2014.01.02,09:06,1.37561,1.37566,1.37558,1.37563,38
2014.01.02,09:07,1.37563,1.37567,1.37561,1.37566,42
2014.01.02,09:08,1.37570,1.37571,1.37564,1.37566,25

我使用以下方式导入它:

raw_data = pd.read_csv('raw_data.csv', engine='c', header=None, index_col=0, names=['date', 'time', 'open', 'high', 'low', 'close', 'volume'], parse_dates=[[0,1]])

但现在我想从数据中提取一些随机(甚至连续)样本，但仅限于连续 5 分钟始终包含数据的样本。因此，例如，无法使用 2014.01.02,08:56 中的数据，因为它存在间隙。但是 2014.01.02,09:00 的数据没问题，因为它在接下来的 5 分钟内始终具有连续的数据。

关于如何有效地实现这一目标有什么建议吗？

Answer 1

这是一种方法，首先 .asfreq('T') 填充一些 NaN，然后使用 rolling_apply 并计算最近是否或者接下来 5 个观察值没有 NaN。

# populate NaNs at minutely freq
# ======================
df = raw_data.asfreq('T')
print(df)

                       open    high     low   close  volume
date_time                                                  
2014-01-02 08:56:00  1.3755  1.3755  1.3755  1.3755       3
2014-01-02 08:57:00     NaN     NaN     NaN     NaN     NaN
2014-01-02 08:58:00     NaN     NaN     NaN     NaN     NaN
2014-01-02 08:59:00     NaN     NaN     NaN     NaN     NaN
2014-01-02 09:00:00  1.3756  1.3756  1.3755  1.3755      21
2014-01-02 09:01:00  1.3755  1.3755  1.3754  1.3755      18
2014-01-02 09:02:00  1.3755  1.3755  1.3755  1.3755      15
2014-01-02 09:03:00  1.3755  1.3756  1.3755  1.3756      39
2014-01-02 09:04:00  1.3756  1.3756  1.3756  1.3756      37
2014-01-02 09:05:00  1.3756  1.3756  1.3756  1.3756      35
2014-01-02 09:06:00  1.3756  1.3757  1.3756  1.3756      38
2014-01-02 09:07:00  1.3756  1.3757  1.3756  1.3757      42
2014-01-02 09:08:00  1.3757  1.3757  1.3756  1.3757      25

consecutive_previous_5min = pd.rolling_apply(df['open'], 5, lambda g: np.isnan(g).any()) == 0
consecutive_previous_5min

date_time
2014-01-02 08:56:00    False
2014-01-02 08:57:00    False
2014-01-02 08:58:00    False
2014-01-02 08:59:00    False
2014-01-02 09:00:00    False
2014-01-02 09:01:00    False
2014-01-02 09:02:00    False
2014-01-02 09:03:00    False
2014-01-02 09:04:00     True
2014-01-02 09:05:00     True
2014-01-02 09:06:00     True
2014-01-02 09:07:00     True
2014-01-02 09:08:00     True
Freq: T, dtype: bool

# use the reverse trick to get the next 5 values
consecutive_next_5min = (pd.rolling_apply(df['open'][::-1], 5, lambda g: np.isnan(g).any()) == 0)[::-1]
consecutive_next_5min

date_time
2014-01-02 08:56:00    False
2014-01-02 08:57:00    False
2014-01-02 08:58:00    False
2014-01-02 08:59:00    False
2014-01-02 09:00:00     True
2014-01-02 09:01:00     True
2014-01-02 09:02:00     True
2014-01-02 09:03:00     True
2014-01-02 09:04:00     True
2014-01-02 09:05:00    False
2014-01-02 09:06:00    False
2014-01-02 09:07:00    False
2014-01-02 09:08:00    False
Freq: T, dtype: bool

# keep rows with either have recent 5 or next 5 elements non-null
df.loc[consecutive_next_5min | consecutive_previous_5min]

                       open    high     low   close  volume
date_time                                                  
2014-01-02 09:00:00  1.3756  1.3756  1.3755  1.3755      21
2014-01-02 09:01:00  1.3755  1.3755  1.3754  1.3755      18
2014-01-02 09:02:00  1.3755  1.3755  1.3755  1.3755      15
2014-01-02 09:03:00  1.3755  1.3756  1.3755  1.3756      39
2014-01-02 09:04:00  1.3756  1.3756  1.3756  1.3756      37
2014-01-02 09:05:00  1.3756  1.3756  1.3756  1.3756      35
2014-01-02 09:06:00  1.3756  1.3757  1.3756  1.3756      38
2014-01-02 09:07:00  1.3756  1.3757  1.3756  1.3757      42
2014-01-02 09:08:00  1.3757  1.3757  1.3756  1.3757      25