CODEWITHSUNDEEP
pythonpandascorrelation
top bantz
top bantz

Reputation: 615

Returning the highest and lowest correlations from a correlation matrix in pandas

I have a bunch of stock data, and I am trying to build a dataframe that takes the top two, and bottom stocks from a correlation matrix, and also their actual correlation.

Let's say the matrix, corr looks like this:

  A    B    C    D    E
A 1.00 0.65 0.31 0.94 0.55
B 0.87 1.00 0.96 0.67 0.41
C 0.95 0.88 1.00 0.72 0.69
D 0.64 0.84 0.99 1.00 0.78
E 0.71 0.62 0.89 0.32 1.00

What I want to do is to be able to is return the best two, and least correlated stocks, and their correlation, for Stock A, B, C, D & E, while dropping the obvious 1.00 correlation each stock has to itself.

The resulting dataframe, or whatever is easiest to display this is would look like this:

Stock 1st 1st_Val 2nd 2nd_Val Last Last_Val
A     D   0.94    B   0.65    C    0.31
B     C   0.96    A   0.87    E    0.41
C     A   0.95    B   0.88    E    0.69
D     C   0.99    B   0.84    A    0.64
E     C   0.89    A   0.71    D    0.32

With my attempts so far I have been able to look through and return the relevant stock names using corr[stock].nlargest().index[0:].tolist(), and then taking [1], [2] and [-1] from each list and sticking them in a dictionary and building the dataframe from there. But I'm unable to return the correlation value and I suspect I'm not doing this in the most efficient way anyway.

Any help really appreciated, cheers

Upvotes: 6

Views: 15999

Answers (4)

Tiago Ferrao
Tiago Ferrao

Reputation: 109

corr.unstack().min() -> to find the value

corr.unstack().idxmin() -> to find the indexes

Upvotes: 1

Adrien Pacifico
Adrien Pacifico

Reputation: 1969

A different answer relying more on the modern pandas style. I did not find a good solution for the second largest correlation. I'll edit the answer when I find it.

### Create an example df
df = pd.DataFrame(data = {"A":pd.np.random.randn(10),
                    "B":pd.np.random.randn(10),
                    "C":pd.np.random.randn(10),
                    "D":pd.np.random.randn(10),
                        }
                )


# Solution
(
df.corr() #correlation matrix
  .replace(1, pd.np.nan) # replace the matrix with nans
  .assign(  # assign new variables
            First = lambda x: x.loc[["A","B","C","D"], ["A","B","C","D"]].idxmax(axis = 1), # Biggest correlation idx
            First_value = lambda x: x.loc[["A","B","C","D"], ["A","B","C","D"]].max(axis = 1), # Biggest correlation
            Last = lambda x: x.loc[["A","B","C","D"],["A","B","C","D"]].idxmin(axis = 1), # Smallest correlation idx
            Last_value = lambda x: x.loc[["A","B","C","D"],["A","B","C","D"]].idxmin(axis = 1), # Smallest correlation
              )
)

I use the .loc[["A","B","C","D"],["A","B","C","D"]] such that the operations are made only on the unmodified data frame.

Output: 
          A         B         C         D First  First_value Last Last_value
A       NaN -0.085776 -0.203110 -0.003450     D    -0.003450    C          C
B -0.085776       NaN -0.110402  0.687283     D     0.687283    C          C
C -0.203110 -0.110402       NaN  0.017644     D     0.017644    A          A
D -0.003450  0.687283  0.017644       NaN     B     0.687283    A          A

Upvotes: 1

CAPSLOCK
CAPSLOCK

Reputation: 6483

If you need to visualize the results but you don't actually need to fetch and work with the actual correlation values, then why not using a very simple heatmap? You could also play with the plot to have the numbers displayed on each square.

import seaborn as sns
import pandas as pd

 dict = {'Date':['2018-01-01','2018-01-02','2018-01-03','2018-01-04','2018-01-05'],'Col1':[1,2,3,4,5],'Col2':[1.1,1.2,1.3,1.4,1.5],'Col3':[0.33,0.98,1.54,0.01,0.99],'Col4':[8,9.98,6,0.01,0.1],'Col1':[19,42,3,0.4,51]}
df = pd.DataFrame(dict, columns=dict.keys())
sns.heatmap(df.corr())

heatmap

Upvotes: 1

pault
pault

Reputation: 43494

Your conditions are hard to generalize into one command, but here is one approach you can take.

Remove the diagonal

import numpy as np
np.fill_diagonal(corr.values, np.nan)
print(corr)
#      A     B     C     D     E
#A   NaN  0.65  0.31  0.94  0.55
#B  0.87   NaN  0.96  0.67  0.41
#C  0.95  0.88   NaN  0.72  0.69
#D  0.64  0.84  0.99   NaN  0.78
#E  0.71  0.62  0.89  0.32   NaN

Find Top 2 and Bottom Column Names

You can use the answer on Find names of top-n highest-value columns in each pandas dataframe row to get the top 2 and bottom one value for each row (Stock).

order_top2 = np.argsort(-corr.values, axis=1)[:, :2]
order_bottom = np.argsort(corr.values, axis=1)[:, :1]

result_top2 = pd.DataFrame(
    corr.columns[order_top2], 
    columns=['1st', '2nd'],
    index=corr.index
)

result_bottom = pd.DataFrame(
    corr.columns[order_bottom], 
    columns=['Last'],
    index=corr.index
)

result = result_top2.join(result_bottom)
#  1st 2nd Last
#A   D   B    C
#B   C   A    E
#C   A   B    E
#D   C   B    A
#E   C   A    D

Now use pandas.DataFrame.lookup to grab the corresponding column value in corr for each column in result

for x in result.columns:
    result[x+"_Val"] = corr.lookup(corr.index, result[x])
print(result)
#  1st 2nd Last  1st_Val  2nd_Val  Last_Val
#A   D   B    C     0.94     0.65      0.31
#B   C   A    E     0.96     0.87      0.41
#C   A   B    E     0.95     0.88      0.69
#D   C   B    A     0.99     0.84      0.64
#E   C   A    D     0.89     0.71      0.32

Reorder columns (optional)

print(result[['1st', '1st_Val', '2nd', '2nd_Val', 'Last', 'Last_Val']])
#  1st  1st_Val 2nd  2nd_Val Last  Last_Val
#A   D     0.94   B     0.65    C      0.31
#B   C     0.96   A     0.87    E      0.41
#C   A     0.95   B     0.88    E      0.69
#D   C     0.99   B     0.84    A      0.64
#E   C     0.89   A     0.71    D      0.32

Upvotes: 6

Related Questions