Level 6 · The correlation heatmap and the strongest pairs

The whole matrix at once

DataFrame.corr() computes r for every pair of numeric columns and returns a square table:

matrix = daily.corr().round(2)
#              spend  sessions  conversions  orders
# spend         1.00      0.71         0.47    0.67
# sessions      0.71      1.00         0.66    0.84
# conversions   0.47      0.66         1.00    0.58
# orders        0.67      0.84         0.58    1.00

Two things are always true: the diagonal is 1 (every column correlates perfectly with itself), and the table is symmetric (spend with sessions equals sessions with spend). Leave identifier columns like order_id out: a correlation with an id number means nothing.

As a heatmap

Sixteen numbers are easier to scan as colors:

ax = sns.heatmap(matrix, annot=True, fmt=".2f", cmap="coolwarm", vmin=-1, vmax=1,
                 linewidths=0.5, square=True)

annot=True writes each value in its cell and fmt=".2f" rounds it. cmap="coolwarm" is a diverging palette: blue for negative, red for positive, pale near zero, which suits a number with a sign. vmin=-1, vmax=1 fixes the scale, so 0.5 always looks like 0.5. Without it, seaborn stretches the colors to your data and a weak 0.4 can look as hot as a strong 0.9.

The strongest pairs, in code

Reading a big matrix by eye misses things, so list the pairs. Each distinct pair is a cell above the diagonal, row i and column j with j > i:

for i in range(len(cols)):
    for j in range(i + 1, len(cols)):
        ...matrix.iloc[i, j]...

Rank them by absolute value, because -0.8 is as strong as +0.8, but keep the sign in the result: it carries the direction. In 2025’s daily numbers, sessions and orders go together most (0.84), followed by sessions and spend (0.71) and orders and spend (0.67). Conversions are only moderately tied to anything, since they’re a small number that jumps around from day to day.

Common mistakes

• Ranking by r instead of abs(r), which pushes strong negative pairs to the bottom.
• Including the diagonal or both halves of the matrix, so every pair shows up twice.
• A heatmap without vmin=-1, vmax=1, whose colors can’t be compared across charts.
• Correlating ids, dates turned into numbers, or other columns that aren’t measurements.
• Reporting only the strongest pairs. A pair close to zero is a finding too: it tells the team which numbers they can stop watching together.