Why Every Machine Learning Engineer Must Learn Excel

Machine Learning · Excel for ML engineers

Why Every ML Engineer Must Learn Excel

SECTION 01Before training: fast data checks in Excel

Before writing any model code, open a sample of the data. Excel lets you scroll, filter and sort in seconds, which often reveals problems that summary statistics hide — wrong units, swapped columns, test rows and impossible values.

import pandas as pd

df = pd.read_parquet("training_data.parquet")
# A random sample is enough to inspect by eye
df.sample(2000, random_state=42).to_excel("sample_check.xlsx", index=False)
# needs: pip install openpyxl

Count blanks in target column C:
=COUNTBLANK(C2:C2001)

Count rows with negative price (column E):
=COUNTIF(E2:E2001,"<0")

Share of each class in column C (label 'yes'):
=COUNTIF(C2:C2001,"yes")/COUNTA(C2:C2001)

SECTION 02After training: error analysis with pivot tables

A single accuracy number hides where a model fails. Export predictions with key features and use pivot tables to see errors by region, product, customer type or month. This tells you what data to collect or which features to improve. It also gives you concrete talking points for team reviews, because you can point to a specific segment, show how many rows it has and explain what you plan to change in the next experiment.

results = X_test.copy()
results["actual"] = y_test.values
results["predicted"] = model.predict(X_test)
results["wrong"] = (results["actual"] != results["predicted"]).astype(int)
results.to_excel("error_analysis.xlsx", index=False)

1. Select the data > Insert > PivotTable
2. Rows: region
3. Values: wrong (Summarize by: Average) -> error rate
4. Add 'Count of wrong' to see sample size
5. Sort by error rate, highest first
6. Ignore groups with very few rows - small samples mislead

SECTION 03Sharing results: speak the business language

Managers, product owners and operations teams usually live in spreadsheets. A clean summary workbook lets them explore results themselves, ask better questions and approve deployment with confidence. Keep the language simple, avoid jargon such as log loss in the summary sheet, and translate every metric into what it means for customers, staff time or revenue.

Model summary - Churn prediction (test period: ______)
Customers scored            : ______
Precision at top 10%        : ______
Recall at top 10%           : ______
Biggest error segment       : ______
Recommended action          : Call top 10% scored customers
Known limits                : New customers have fewer signals

with pd.ExcelWriter("model_report.xlsx") as writer:
    summary.to_excel(writer, sheet_name="Summary", index=False)
    by_region.to_excel(writer, sheet_name="Errors by region")
    review_list.to_excel(writer, sheet_name="Review list", index=False)

SECTION 04Templates: confusion matrix and review sheet

Build these once and reuse them in every classification project.

Data: column A = actual, column B = predicted (rows 2-1001)
Layout:            Pred: yes        Pred: no
Actual: yes        =COUNTIFS($A$2:$A$1001,"yes",$B$2:$B$1001,"yes")   =COUNTIFS($A$2:$A$1001,"yes",$B$2:$B$1001,"no")
Actual: no         =COUNTIFS($A$2:$A$1001,"no",$B$2:$B$1001,"yes")    =COUNTIFS($A$2:$A$1001,"no",$B$2:$B$1001,"no")
Precision = TP / (TP + FP)
Recall    = TP / (TP + FN)

Columns for a team labelling review:
ID | Text or image link | Model label | Reviewer label | Agree? | Notes
Agree? formula (row 2): =IF(C2=D2,"Yes","No")
Agreement rate: =COUNTIF(E2:E501,"Yes")/COUNTA(E2:E501)
Use Data Validation lists for Reviewer label to avoid typos.

# Read reviewer labels back into Python
reviewed = pd.read_excel("labelling_review.xlsx")
reviewed = reviewed[reviewed["Reviewer label"].notna()]
train_df = train_df.merge(reviewed[["ID", "Reviewer label"]], on="ID", how="left")
train_df["label"] = train_df["Reviewer label"].fillna(train_df["label"])

SECTION 05Tips for ML engineers using Excel

Use Excel where it is strong, and Python where it is strong:

• Keep Python as the source of truth — Excel is for inspection and sharing; pipelines stay in code.
• Export samples, not everything — large files are slow in Excel and add little value.
• Use XLOOKUP and COUNTIFS fluently — they cover most quick analysis needs.
• Protect sensitive data — mask personal information before sharing workbooks.
• Name sheets clearly — Summary, Errors, Review list — readers should not have to guess.

SECTION 06Interview Q&A — Excel for ML engineers

SECTION 07Test yourself — Excel for ML engineers quiz

SECTION 08Frequently asked questions

SECTION 09Related reads