A few months back, I sat down with a dataset from a fictional-but-very-real-feeling company called Austo Motor Company — a manufacturer of SUVs, Sedans, and Hatchbacks. The brief was simple on paper: the board wasn’t happy with how the marketing campaign was performing, and they wanted a data scientist to figure out why, and more importantly, who they should actually be targeting.
I’ve done a fair number of exploratory data analysis projects by now, but this one stuck with me because it followed a pattern I keep running into in real business analytics work: the data is messier than the brief suggests, the “obvious” assumptions turn out to be wrong, and the real insight is usually hiding one groupby away from where you first looked. This post walks through exactly how I approached it — cleaning, exploring, questioning, and finally turning raw customer rows into a targeting strategy the marketing team could actually act on.
If you want to follow along with the actual code, the full notebook and dataset structure are on my GitHub as part of the Automobile Sales Data Insights project.
The business problem: a marketing budget going somewhere, but nobody's sure where
Here’s the setup. Austo Motor Company sells three types of vehicles — SUVs, Sedans, and Hatchbacks — and their board had raised a concern that’s painfully common in mid-size companies: we’re spending on marketing, but we’re not confident it’s reaching the right people.
My job, as the analyst brought in, was to dig into customer demographic and financial data and answer a very grounded set of questions:
- Who is actually buying which car type — and does it match who marketing is targeting?
- Are there financial or demographic signals (income, loans, marital status, profession) that predict a purchase?
- Where is the campaign budget likely being wasted on the wrong audience?
Notice what’s not in that list — there’s no ask for a fancy model, no request for a dashboard. Just: understand the customer, and tell us how to spend smarter. That’s a good reminder that most business analytics work starts and ends with clean EDA, not machine learning.
Meeting the dataset
The dataset had 1,581 rows and 14 columns of customer demographic and financial information — a reasonably sized real-world sample. Once I ran df.info(), the structure became clear: 6 numerical columns and 8 categorical (object-type) columns, covering things like:
| Category | Columns |
|---|---|
|
Demographics |
Age, Gender, Marital Status, Education, Profession |
|
Household |
No. of Dependents, Partner Working |
|
Finances |
Salary, Partner Salary, Total Salary, Personal Loan, House Loan |
|
Purchase |
Make (car type), Price |
Right away, df.info() flagged something worth chasing: the Gender and Partner_salary columns had fewer non-null entries than the rest — a classic sign of missing data hiding in plain sight. Before touching a single chart, I knew the first job was cleanup.
What the numbers looked like at a glance
Age ranged from 22 to 54, with half the customers under 29. Salaries ran from ₹30,000 to ₹99,300 (average ≈ ₹60,392), and total household income (salary + partner salary) stretched up to ₹1,71,000. Car prices ranged from ₹18,000 to ₹70,000. Nothing shocking yet — but the shape of these distributions would matter a lot later.
Cleaning the mess: missing values and typos
This is the part of every project that nobody puts on their portfolio slide, and it’s usually where the real thinking happens. Running df.isnull().sum() confirmed it: 159 missing values in total — 53 in Gender and 106 in Partner_salary.
For Gender, there wasn’t a mathematically “correct” way to guess someone’s gender from the rest of their row, so I made the honest call: fill it with "Unknown" rather than force an assumption I couldn’t justify. Silently guessing here would have quietly biased every gender-based insight later in the project — and that’s exactly the kind of hidden decision that sinks the credibility of an analysis if the client discovers it later.
Partner_salary was more interesting, because the data itself gave me a way to solve it logically. If a customer’s partner wasn’t working, it made sense that their partner salary was genuinely zero — not missing, just absent. But if the partner was working and the salary was blank, I could reconstruct it using a relationship that already existed in the data: Total_Salary − Salary = Partner_Salary.
def fill_partner_salary(row):
# If the partner isn't working, their income is genuinely 0
if row['Partner_working'] == 'No':
return 0
# Otherwise, back it out from the total household income
else:
return row['Total_salary'] - row['Salary']
df['Partner_salary'].fillna(df.apply(fill_partner_salary, axis=1), inplace=True)
df['Gender'].fillna('Unknown', inplace=True)This is a small snippet, but it’s the piece of the project I’m most proud of — it’s the difference between “filling blanks” and actually understanding the business logic well enough to recover real values instead of guessing.
Next came the classic categorical audit — printing .unique() on every object column to catch inconsistent labels before they quietly wreck a groupby. Sure enough, the Gender column had two misspelled variants, “Femal” and “Femle”, sitting alongside “Female.” I corrected both with a simple .replace(), then re-ran the unique-value check across every categorical column to confirm the dataset was finally clean and analysis-ready.
⚠️ Why this step matters more than people think
If I’d skipped the typo check, “Femal” and “Femle” would have shown up as two separate, tiny categories in every downstream chart and groupby — quietly under-counting the real female segment. Small spelling errors like this are one of the most common (and most invisible) sources of wrong business conclusions.
Univariate EDA: getting to know each column on its own
With the data clean, I moved into univariate analysis — looking at one variable at a time before mixing them together. I split this into categorical and numerical passes.
A few patterns jumped out immediately:
- Male customers significantly outnumbered female customers in the dataset.
- Salaried professionals bought more cars than business owners.
- Married individuals were far more likely to be buyers than single individuals.
- Post-graduates outnumbered graduates among buyers.
- Customers without a house loan bought more cars than those carrying one.
- Personal loan status barely mattered — buyers with and without one were roughly split evenly.
- Sedans were, by far, the most purchased car type overall.
For Age, Salary, Partner Salary, Total Salary, and Price, I plotted histograms with KDE curves and boxplots side by side. Every single one of these distributions was right-skewed — a long tail of higher earners and pricier purchases, with most customers clustered toward the lower-to-mid range.
The boxplots flagged outliers in Total Salary and No. of Dependents. This is where I made a judgment call I’d make again: I decided not to remove them. A household with zero dependents or a total income above ₹1,40,000 isn’t a data-entry error — it’s a real customer. Stripping them out would have quietly deleted the company’s highest-value segment right before I needed to analyze high-value segments. Statistical outliers and business-invalid values are not the same thing, and mixing them up is one of the most damaging mistakes an analyst can make.
I worked in Google Colab, mostly because I didn’t want to fight with a local environment on a Saturday afternoon. Mounting Drive and loading the CSV took two lines:
Bivariate EDA: how age, salary, and price move together
Next, I looked at how the numerical variables related to each other using a pairplot and a correlation heatmap across Age, No. of Dependents, Salary, Partner Salary, Total Salary, and Price.
🔍 What the correlations revealed
Age and Price were strongly correlated — as customers got older, they tended to buy pricier cars. Meanwhile, No. of Dependents had a negative correlation with Age, Salary, and Price — households with more dependents leaned toward younger buyers with lower income and cheaper cars.
This single correlation pair — age and price moving together — quietly set up the entire targeting strategy I’d land on later: older, more established customers are the ones ready for a premium purchase, while younger buyers with more financial responsibilities need a lower entry point.
The six questions the board actually wanted answered
Here’s where the project stopped being “just EDA” and turned into genuine business analysis. The Data Science team had six specific, pointed questions. I’ll walk through what I found for each one — including a couple of results that flipped my initial assumptions.
My gut assumption going in was “probably yes.” The data disagreed. When I grouped by Make and looked at gender distribution:
| Car Type | % Male Buyers | % Female Buyers |
|---|---|---|
|
Hatchback |
97.08% |
2.58% |
|
Sedan |
73.50% |
20.09% |
|
SUV |
39.73% |
58.25% |
🔄 The twist
Women were actually the majority buyers of SUVs (58.25%), while men overwhelmingly dominated Sedans and Hatchbacks. The real story wasn’t “men prefer SUVs” — it was the opposite: relative to their overall smaller share of buyers, women skew heavily toward SUVs, while men are spread across all three types with Sedan as their clear favorite.
Grouping by Profession and Make told a consistent story: salaried professionals out-purchased business owners across every single car type, not just Sedans.
| Car Type | % Salaried Buyers | % Business Buyers |
|---|---|---|
|
Hatchback |
50.17% |
49.83% |
|
Sedan |
56.41% |
43.59% |
|
SUV |
70.00% |
30.00% |
So a salaried person buying a Sedan isn’t just likely — it’s the single most common purchase pattern in the entire dataset. And interestingly, salaried buyers were most concentrated in SUVs (70%), not Sedans, even though Sedans win on total volume.
This question was framed around a claim — the kind of confident, “obviously true” statement that shows up in real boardrooms all the time. I sliced the data by Profession and Gender together to actually test it.
The result: salaried males were the least likely to be SUV buyers among all the segments — they overwhelmingly preferred Sedans and Hatchbacks. The claim didn’t hold up.
📊 What the data actually showed
- Salaried males strongly prefer Sedans (43.59% of their purchases) — not SUVs.
- Business males had the lowest SUV preference of any group (just 5.9%).
- Salaried females were the segment genuinely easiest to target for SUVs.
- Business females also leaned slightly toward SUVs over Sedans.
- Hatchbacks were most popular with business males specifically.
This is exactly the kind of assumption-testing that makes EDA worth doing — a confident statement from someone in the room turned out to be backwards once I checked it against real purchase behavior.
I compared average car price by gender directly:
| Gender | Average Price Paid |
|---|---|
|
Female |
₹47,705 |
|
Male |
₹32,817 |
Women spent, on average, about 45% more per car than men. Combined with the earlier finding that women skew heavily toward SUVs (typically the pricier segment), this made complete sense — and it directly contradicted any assumption that women are a lower-value or price-sensitive segment.
| Personal Loan | Average Price Paid |
|---|---|
|
Yes |
₹34,457 |
|
No |
₹36,742 |
The difference here was small — only around 6%. Having an existing personal loan doesn’t meaningfully change how much someone spends on a car. It’s a variable I initially expected to matter more than it did.
Customers whose partner was not working actually spent slightly more on average than those with a working partner — the opposite of what I expected going in. But the gap was small enough that I wouldn’t call it a strong driver of purchase price on its own.
Turning insights into a targeting strategy
Data without a decision attached to it is just trivia. Here’s how I translated the six answers above into something the marketing team could actually run with:
✅ Business recommendations
- Re-target SUV campaigns toward women and salaried female professionals — the current messaging likely skews male by default, but the real high-intent SUV buyer is a working woman with strong purchasing power.
- Lean into Sedans for salaried male professionals — this is the single largest, most reliable segment in the entire dataset, and marketing spend here has the most predictable return.
- Position Hatchbacks as the entry-level “first car” for business-professional males and younger, lower-income households with more dependents — this segment is price-sensitive, not brand-sensitive.
- Prioritize married customers without an existing house loan as the core acquisition segment — they showed up as buyers far more often than singles or customers already carrying housing debt.
- Don’t over-invest in loan-based messaging — personal loan status barely moved the needle on spend, so “financing available” as a headline benefit isn’t the persuasion lever it might seem to be.
- Use age as a soft signal for premium upsell — since age correlates strongly with price, older customers already in the funnel are good candidates for a premium trim or SUV upsell conversation.
What this project taught me
A few things stuck with me well after I closed this notebook:
- Segmentation isn’t just about who buys — it’s about who buys what. Slicing by profession and gender together (not separately) is what actually busted the Sheldon Cooper assumption. Single-variable analysis would have missed it completely.
- Data storytelling beats data reporting. “56.41% of Sedan buyers are salaried” is a statistic. “Salaried professionals are your most predictable Sedan segment, and your current SUV messaging is probably aimed at the wrong gender” is a decision.
- Feature logic beats blind imputation. Reconstructing Partner_salary from Total_salary − Salary instead of using a mean or median fill preserved the actual relationship in the data, instead of introducing noise.
- Outliers aren’t automatically errors. Knowing when to leave them in mattered as much as knowing how to detect them.
What I'd do next
This project answered the board’s immediate questions, but it also opened up a clear roadmap if I picked it back up:
- Predictive modeling — a classification model to predict the likely car type for a new lead before they even talk to a salesperson.
- A live dashboard — porting these segment breakdowns into Power BI or Tableau so the marketing team can filter by segment themselves instead of waiting on a notebook.
- External macroeconomic data — layering in inflation or interest-rate trends to see how they shift the Sedan-vs-SUV-vs-Hatchback mix over time.
Key takeaways
The biggest wins in this project didn’t come from a clever chart — they came from questioning assumptions that everyone in the room already believed were true.
- 159 missing values and two misspelled gender labels were quietly sitting between the raw data and a trustworthy analysis.
- The “men prefer SUVs” and “salaried males are easy SUV targets” assumptions were both wrong — women and salaried females were the real SUV segment.
- Women outspent men by ~45% on average per car — a segment worth far more marketing attention than it was likely getting.
- Married customers without a house loan, and salaried professionals overall, are the highest-confidence segments to target across all three car types.
Want the full notebook and code ?
The complete data cleaning steps, EDA visualizations, and analysis are on my GitHub, along with more end-to-end data science case studies.
Add comment
You must be logged in to post a comment.