How to manipulate data cell by cell in pandas df?

Question

Let the sample df (df1) be,

We can achieve df2 or final data-frame by manipulating the data of df1 in the following manner,

Step 1: Remove all positive numbers including zeros

After Step 1 the sample data should look like,

Step 2: If column4 A row is a negative number and column4 B is blank, then remove the -ve number of column4 A row

Step 3: If column4 A row is blank and column4 B is a negative number, then keep the -ve number of column4 B row

After Steps 1,2 and 3 are done,

Step 4: If both A and B of column4 are negative then,

For each A and B row of column4, check the left-side (LHS) value (for a given month) of the same A and B row of column4

Step 4.1: If either of the LHS values of A or B is a -ve number, then delete the current row value of B column4 and keep the current row value of A column4

After Step 4.1, the sample data should look like this,

Step 4.2:

If the LHS value of A and B column4 is blank, then keep the current row value of B column4 and delete the current row value of A column4

Sample data after Step 4.2 should look like,

Since we see two negative numbers still, we perform Step 4.1 again and then the final data-frame or df2 will look like,

How may I achieve the above using pandas? I was able to achieve till Step 1 but have no idea as to how to proceed further. Any help would be greatly appreciated.

This is the approach that I took,

import pandas as pd
df = pd.read_excel('df1.xlsx', engine='openpyxl')
df.to_pickle('./df1.pkl')
unpickled_df = pd.read_pickle('./df1.pkl')
rem_cols = ['column2', 'column3', 'column5', 'column6', 'column7']
unpickled_df['g'] = unpickled_df.groupby(['column1', 'column4'] ).cumcount()
df1 = unpickled_df.drop(rem_cols, axis=1)
df1 = df1.set_index(['column1','g', 'column4'])
df1.columns = pd.to_datetime(df1.columns, format='%b-%y').strftime('%b-%y')
first_date = df1.columns[0]
df1 = df1.unstack(-1)
df1 = df1.mask(df1.ge(0))
m1 = (df1.xs('A', level=1, axis=1, drop_level=False).notna() & 
      df1.xs('B', level=1, axis=1, drop_level=False).rename(columns={'B':'A'}, level=1).isna())
m2 = (df1.xs('B', level=1, axis=1, drop_level=False).notna() &
      df1.xs('A', level=1, axis=1, drop_level=False).rename(columns={'A':'B'}, level=1).isna())

m = m1.join(m2)
df1 = df1.mask(m)
df2 = df1.groupby(level=1, axis=1).shift(1, axis=1)
mask1 = df1.notna() & df2.isna() & (df1.columns.get_level_values(1) == 'A')[ None, :]
mask1[first_date] = False
mask2 = df1.notna() & df2.notna() & (df1.columns.get_level_values(1) == 'B')[ None, :]
df1 = df1.mask(mask1).mask(mask2).stack(dropna=False)
unpickled_df = unpickled_df[rem_cols + ['column1','g', 'column4']].join(df1, on=['column1','g', 'column4'])
#print(unpickled_df)

Small Test data: df1,

{'column1': ['ABC', 'ABC', 'CDF', 'CDF'], 'column4': ['A', 'B', 'A', 'B'], 'Feb-21': [0, 10, 0, 0], 'Mar-21': [0, 0, 70, 70], 'Apr-21': [-10, -10, -8, 60], 'May-21': [-30, -60, -10, 40], 'Jun-21': [-20, 9, -40, -20], 'Jul-21': [30, -10, 0, -20], 'Aug-21': [-30, -20, 0, -20], 'Sep-21': [0, -15, 0, -20], 'Oct-21': [0, -15, 0, -20]}

df2 (expected output),

{'column1': ['ABC', 'ABC', 'CDF', 'CDF'], 'column4': ['A', 'B', 'A', 'B'], 'Feb-21': [nan, nan, nan, nan], 'Mar-21': [nan, nan, nan, nan], 'Apr-21': [nan, -10.0, nan, nan], 'May-21': [-30.0, nan, nan, nan], 'Jun-21': [nan, nan, nan, -20.0], 'Jul-21': [nan, -10.0, nan, -20.0], 'Aug-21': [-30.0, nan, nan, -20.0], 'Sep-21': [nan, -15.0, nan, -20.0], 'Oct-21': [nan, -15.0, nan, -20.0]}

Test data:

df1

{'column1': ['CT', 'CT', 'NBB', 'NBB', 'CT', 'CT', 'NBB', 'NBB', 'HHH', 'HHH', 'TP1', 'TP1', 'TPR', 'TPR', 'PP1', 'PP1', 'PP1', 'PP1'], 'column2': ['POUPOU', 'POUPOU', 'PRPRP', 'PRPRP', 'STDD', 'STDD', 'STDD', 'STDD', 'STEVT', 'STEVT', 'SYSYS', 'SYSYS', 'SYSYS', 'SYSYS', 'SHW', 'SHW', 'JV', 'JV'], 'column3': ['V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV'], 'column4': ['A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B'], 'column5': [nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan], 'column6': ['BBB', 'BBB', 'CCC', 'CCC', 'BBB', 'BBB', 'BBB', 'BBB', 'VVV', 'VVV', 'CHCH', 'CHCH', 'CHCH', 'CHCH', 'CCC', 'CCC', 'CHCH', 'CHCH'], 'column7': ['Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Mar-21', 'Mar-21', 'Mar-21', 'Mar-21', 'Mar-21', 'Mar-21', 'Apr-21', 'Apr-21', 'Mar-21', 'Mar-21'], 'Feb-21': [11655, 0, 0, 0, 121117, 0, 14948, 0, 0, 0, 0, 0, 0, 0, 1838, 0, 0, 0], 'Mar-21': [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -16474.0, -16474.0, 7000.0, 7000.0, -19946.0, -19946.0, 16084.44444444444, 0.0, 0.0, 0.0], 'Apr-21': [104815.0, 104815.0, 17949.0, 17949.0, 96132.0, 96132.0, 0.0, 0.0, -17001.0, -33475.0, -878.0, 6122.0, 8398.0, -11548.0, -5297.073170731703, -5297.073170731703, -282.0, -282.0], 'May-21': [78260.0, 183075.0, 42557.0, 60506.0, -15265.0, 80867.0, -18.0, -18.0, 21084.0, -12391.0, -1831.0, 4291.0, 2862.0, -8686.0, 5261.25, -35.8231707317027, -369.0, -651.0], 'Jun-21': [-52480.0, 130595.0, -13258.0, 47248.0, -35577.0, 45290.0, 2434.0, 2416.0, 31147.0, 18756.0, -4310.0, -19.0, -4750.0, -13436.0, -92.0, -127.8231707317027, -280.0, -931.0], 'Jul-21': [-174544.0, -43949.0, -38127.0, 9121.0, -124986.0, -79696.0, -9707.0, -7291.0, 13577.0, 32333.0, 0.0, -19.0, -15746.0, -29182.0, 93.0, -34.8231707317027, -319.0, -1250.0], 'Aug-21': [35498.0, -8451.0, -37094.0, -27973.0, 79021.0, -675.0, -1423.0, -8714.0, 32168.0, 64501.0, 0.0, -19.0, 18702.0, -10480.0, 4347.634146341465, 4312.810975609762, -341.0, -1591.0], 'Sep-21': [44195.0, 35744.0, 2039.0, -25934.0, 70959.0, 70284.0, 2816.0, -5898.0, 38359.0, 102860.0, 0.0, -19.0, 18119.0, 7639.0, 5302.222222222219, 9615.033197831981, 0.0, -1591.0], 'Oct-21': [-13163.0, 22581.0, -4773.0, -30707.0, 205080.0, 275364.0, -709.0, -6607.0, -1397.0, 101463.0, 0.0, -19.0, 0.0, 7639.0, -34.0, 9581.033197831981, 0.0, -1591.0]}

df2 (expected output) ,

{'column1': ['CT', 'CT', 'NBB', 'NBB', 'CT', 'CT', 'NBB', 'NBB', 'HHH', 'HHH', 'TP1', 'TP1', 'TPR', 'TPR', 'PP1', 'PP1', 'PP1', 'PP1'], 'column2': ['POUPOU', 'POUPOU', 'PRPRP', 'PRPRP', 'STDD', 'STDD', 'STDD', 'STDD', 'STEVT', 'STEVT', 'SYSYS', 'SYSYS', 'SYSYS', 'SYSYS', 'SHW', 'SHW', 'JV', 'JV'], 'column3': ['V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV', 'V', 'CV'], 'column4': ['A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B'], 'column5': [nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan], 'column6': ['BBB', 'BBB', 'CCC', 'CCC', 'BBB', 'BBB', 'BBB', 'BBB', 'VVV', 'VVV', 'CHCH', 'CHCH', 'CHCH', 'CHCH', 'CCC', 'CCC', 'CHCH', 'CHCH'], 'column7': ['Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Apr-21', 'Mar-21', 'Mar-21', 'Mar-21', 'Mar-21', 'Mar-21', 'Mar-21', 'Apr-21', 'Apr-21', 'Mar-21', 'Mar-21'], 'Feb-21': [nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan], 'Mar-21': [nan, nan, nan, nan, nan, nan, nan, nan, nan, -16474.0, nan, nan, nan, -19946.0, nan, nan, nan, nan], 'Apr-21': [nan, nan, nan, nan, nan, nan, nan, nan, -17001.0, nan, nan, nan, nan, -11548.0, nan, -5297.073170731703, nan, -282.0], 'May-21': [nan, nan, nan, nan, nan, nan, nan, -18.0, nan, -12391.0, nan, nan, nan, -8686.0, nan, -35.8231707317027, -369.0, nan], 'Jun-21': [nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, -19.0, -4750.0, nan, -92.0, nan, -280.0, nan], 'Jul-21': [nan, -43949.0, nan, nan, nan, -79696.0, nan, -7291.0, nan, nan, nan, -19.0, -15746.0, nan, nan, -34.8231707317027, -319.0, nan], 'Aug-21': [nan, -8451.0, nan, -27973.0, nan, -675.0, -1423.0, nan, nan, nan, nan, -19.0, nan, -10480.0, nan, nan, -341.0, nan], 'Sep-21': [nan, nan, nan, -25934.0, nan, nan, nan, -5898.0, nan, nan, nan, -19.0, nan, nan, nan, nan, nan, -1591.0], 'Oct-21': [nan, nan, -4773.0, nan, nan, nan, -709.0, nan, nan, nan, nan, -19.0, nan, nan, nan, nan, nan, -1591.0]}

Note: I have implemented my code on the basis of the Test data provided. The sample data is merely to focus on the columns that are supposed to be manipulated.

Answer 1

What about this ?

For each step, I group on column1, then set column4 as index and work on the transpose matrix with your criteria. Note that I've extrapoled a bit on your criteria to match your attended results (I hope it is correct but you will have to check that).

Note also that I have kept each step separate to make it easier to read. But it would be more efficient to make the grouping/indexing/transposing in one shot and work on your algorithm from there.

import pandas as pd
import numpy as np
pd.set_option('display.max_columns', None)


df1 = pd.DataFrame(
        {'column1': ['ABC', 'ABC', 'CDF', 'CDF'], 'column4': ['A', 'B', 'A', 'B'], 'Feb-21': [0, 10, 0, 0], 'Mar-21': [0, 0, 70, 70], 'Apr-21': [-10, -10, -8, 60], 'May-21': [-30, -60, -10, 40], 'Jun-21': [-20, 9, -40, -20], 'Jul-21': [30, -10, 0, -20], 'Aug-21': [-30, -20, 0, -20], 'Sep-21': [0, -15, 0, -20], 'Oct-21': [0, -15, 0, -20]}
        )

print(df1)
print('-'*50)

df = df1.copy()

#step1 :
df.iloc[:, 2:] = df.iloc[:, 2:].where(df.iloc[:, 2:] < 0, np.nan)


print(df)
print('-'*50)

#step2 :
def step2(df):
    df = df.set_index("column4").T
    ix = df[(df.A<=0) & (df.B.isnull())].index
    df.loc[ix, "A"] = np.nan
    return df.T
df = df.groupby('column1').apply(step2)
df.reset_index(drop=False, inplace=True)
print(df)
print('-'*50)


#step3 :
def step3(df):
    df = df.set_index("column4").T
    ix = df[(df.A.isnull()) & (df.B>=0)].index
    df.loc[ix, "B"] = df.loc[ix, "A"]
    return df.T
df = df.groupby('column1').apply(step3)
df.reset_index(drop=False, inplace=True)
print(df)
print('-'*50)

#step4 :
def step4(df):
    df = df.set_index("column4").T
    a_pos = df.columns.get_loc('A')
    b_pos = df.columns.get_loc('B')
    #step 4.1
    ix = df[(df.A<0) & (df.B<0)].index
    if len(ix):
        ix = df.index.get_indexer(ix)
        left_pos = ix-1
        condition_left = df.iloc[left_pos].notnull().any(axis=1)
        ix = condition_left[condition_left].index
        ix = df.index.get_indexer(ix)
        df.iloc[ix+1, b_pos] = np.nan
    
    #step 4.2
    ix = df[(df.A<0) & (df.B<0)].index
    if len(ix):
        ix = df.index.get_indexer(ix)
        left_pos = ix-1
        condition_left = df.iloc[left_pos].isnull().all(axis=1)    
        ix = condition_left[condition_left].index
        ix = df.index.get_indexer(ix)
        df.iloc[ix+1, a_pos] = np.nan
    
    #step 4.1 (again)
    ix = df[(df.A<0) & (df.B<0)].index
    if len(ix):
        ix = df.index.get_indexer(ix)
        left_pos = ix-1
        condition_left = df.iloc[left_pos].notnull().any(axis=1)
        ix = condition_left[condition_left].index
        ix = df.index.get_indexer(ix)
        df.iloc[ix+1, b_pos] = np.nan
        
    return df.T

df = df.groupby('column1').apply(step4)
df.reset_index(drop=False, inplace=True)
print(df)
print('-'*50)

EDIT

I'll assume here (based on your previous comment) that your dataframe will always be composed of A/B rows in alternance (and that the order in the dataframe is valid). We will then need to compute an artificial index to indentify each pair of rows.

Note that I used a fillna on your columns (mainly column5) as it is good practice with the groupby commands. Due to multiple levels of columns, I'm not sure it will have an impact anyway...

Boolean indexing begins to be tricky when managing multiple levels of columns. You will see I will compute each "column" to a numpy.array (using the .values method). Somehow, pandas won't perform the boolean match on multiple columns, I'm not exactly sure why.

So this goes :

df = pd.DataFrame(  ...  )

#Compute the unique index for each pair of rows
df.reset_index(drop=False, inplace=True)
ix = df.index
ix = ix[ix%2==0]
df.loc[ix, 'index'] = df.shift(-1).loc[ix, 'index']

#step1 :
cols = [x for x in df.columns.tolist() if not x.startswith('column') and x != "index"]
df[cols] = df[cols].where(df[cols] < 0, np.nan)


cols_index = ["column4", "column1", "column2", "column3", "column5", "column6", "column7"]
df[cols_index] = df[cols_index].fillna(-1)

#step2 :
def step2(df):
    df = df.set_index(cols_index).drop('index', axis=1).T
    ix = df[
            (df.A<=0).values
            & df.B.isnull().values
         ].index
    df.loc[ix, "A"] = np.nan
    return df.T
df = df.groupby('index').apply(step2)
print(df)
df.reset_index(drop=False, inplace=True)
print(df)
print('-'*50)


#step3 :
def step3(df):
    df = df.set_index(cols_index).drop('index', axis=1).T
    ix = df[
            df.A.isnull().values 
            & (df.B>=0).values
            ].index
    df.loc[ix, "B"] = df.loc[ix, "A"]
    return df.T
df = df.groupby('index').apply(step3)
df.reset_index(drop=False, inplace=True)
print(df)
print('-'*50)

#step4 :
def step4(df):
    df = df.set_index(cols_index).drop('index', axis=1).T
    a_pos = df.columns.get_loc('A')
    b_pos = df.columns.get_loc('B')
    
    #step 4.1
    ix = df[
            (df.A<0).values
            & (df.B<0).values
            ].index
    if len(ix):
        ix = df.index.get_indexer(ix)
        left_pos = ix-1
        condition_left = df.iloc[left_pos].notnull().any(axis=1)
        ix = condition_left[condition_left].index
        ix = df.index.get_indexer(ix)
        df.iloc[ix+1, b_pos] = np.nan
    
    #step 4.2
    ix = df[
            (df.A<0).values
            & (df.B<0).values
            ].index
    if len(ix):
        ix = df.index.get_indexer(ix)
        left_pos = ix-1
        condition_left = df.iloc[left_pos].isnull().all(axis=1)    
        ix = condition_left[condition_left].index
        ix = df.index.get_indexer(ix)
        df.iloc[ix+1, a_pos] = np.nan
    
    #step 4.1 (again)
    ix = df[
            (df.A<0).values
            & (df.B<0).values
            ].index
    if len(ix):
        ix = df.index.get_indexer(ix)
        left_pos = ix-1
        condition_left = df.iloc[left_pos].notnull().any(axis=1)
        ix = condition_left[condition_left].index
        ix = df.index.get_indexer(ix)
        df.iloc[ix+1, b_pos] = np.nan
        
    return df.T

df = df.groupby('index').apply(step4)
df.reset_index(drop=False, inplace=True)
print(df)
print('-'*50)

And if you want to restore your column5 :

df[cols_index] = df[cols_index].replace(-1, np.nan)