Python 2.7 Pandas: How to replace a for-loop? - python-2.7

Question

I have a large pandas dataframe with 2000 rows (one date per row) and 2000 columns (1 second intervals). Each cell represents a temperature reading.
Starting with the 5th row, I need to go back 5 rows and find all the observations where the the 1st column in the row is higher than the 2nd column in the row.
For the 5th row I may find 2 such observations. I then want to do summary stats on the observations and append those summary stats to a list.
Then I go to the 6st row and go back 5 rows and find all the obvs where the 1th column is higher than the 2nd column. I get all obvs, do summary stats on the obvs and append the results to the new dataframe.
So for each row in the dataframe, I want to go back 5 days, get the obvs, get the stats, and append the stats to a dataframe.
The problem is that if I perform this operation on rows 5 -2000, then I will have a for-loop that is 1995 cycles long, and this takes a while.
What is the better or best way to do this?
Here is the code:
print huge_dataframe
sec_1 sec_2 sec_3 sec_4 sec_5
2013_12_27 0.05 0.12 0.06 0.15 0.14
2013_12_28 0.06 0.32 0.56 0.14 0.17
2013_12_29 0.07 0.52 0.36 0.13 0.13
2013_12_30 0.02 0.12 0.16 0.55 0.12
2013_12_31 0.06 0.30 0.06 0.14 0.01
2014_01_01 0.05 0.12 0.06 0.15 0.14
2014_01_02 0.06 0.32 0.56 0.14 0.17
2014_01_03 0.07 0.52 0.36 0.13 0.13
2014_01_04 0.02 0.12 0.16 0.55 0.12
2014_01_05 0.06 0.30 0.06 0.14 0.01
for each row in huge_dataframe.ix[5:]:
move = row[sec_1] - row[sec_2]
if move < 0: move = 'DOWN'
elif move > 0: move = 'UP'
relevant_dataframe = huge_dataframe.ix[only the 5 rows preceding the current row]
if move == 'UP':
mask = relevant_dataframe[sec_1 < sec_2] # creates a boolean dataframe
observations_df = relevant_dataframe[mask]
elif move == 'DOWN':
mask = relevant_dataframe[sec_1 > sec_2] # creates a boolean dataframe
observations_df = relevant_dataframe[mask]
# At this point I have observations_df which is only filled
# with rows where sec_1 < sec_2 or the opposite, depending on which
# row I am in.
summary_stats = str(observations_df.describe())
summary_list.append(summary_stats) # This is the goal
# I want to ultimatly
# turn the list into a
# dataframe

Answer 1

Since there is no code to create the data, I will just sketch the code that I would try to make work. Generally, try to prevent from row-wise operations whenever you can. I first had no clue either, but then I got interested and some research yielded TimeGrouper:
df = big_dataframe
df['move'] = df['sec_1'] > df['sec2']
def foobarRules(group):
# keep in mind that in here, you refer not to "relevant_dataframe", but to "group"
if (group.tail(1).move == True):
# some logic
else:
# some other logic
return str(group.describe())
grouper = TimeGrouper('5D')
allMyStatistics = df.groupby(grouper).apply(foobarRules)
I have honestly no clue how the return works if you return a multi-dimensional dataframe. I know it works well if you return either a row or a column, but if you return a dataframe that contains both rows and columns for every group - I guess pandas is smart enough to compute a panel of all these. Well, you will find out.