EDA in Public (Half 3): RFM Evaluation for Buyer Segmentation in Pandas

! If you happen to’ve been following alongside, we’ve come a great distance. In Half 1, we did the “soiled work” of cleansing and prepping.

In Half 2, we zoomed out to a high-altitude view of NovaShop’s world — recognizing the large storms (high-revenue nations) and the seasonal patterns (the huge This autumn rush).

However right here’s the factor: a enterprise doesn’t truly promote to “months” or “nations.” It sells to human beings.

If you happen to deal with each buyer precisely the identical, you’re making two very costly errors:

Over-discounting: Giving a “20% off” coupon to somebody who was already reaching for his or her pockets.
Ignoring the “Quiet” Ones: Failing to note when a previously loyal buyer stops visiting, till they’ve been gone for six months and it’s too late to win them again.

The Answer? Behavioural Segmentation.

As an alternative of guessing, we’re going to make use of the information to let the shoppers inform us who they’re. We do that utilizing the gold customary of retail analytics: RFM Evaluation.

Recency (R): How not too long ago did they purchase? (Are they nonetheless engaged with us?)
Frequency (F): How usually do they purchase? (Are they loyal, or was it a one-off?)
Financial (M): How a lot do they spend? (What’s their whole enterprise influence?)

By the top of this half, we’ll transfer past “Prime 10 Merchandise” and really assign a particular, actionable Label to each single buyer in NovaShop’s database.

Information Preparation: The “Lacking ID” Pivot

Earlier than we are able to begin scoring, we’ve got to deal with a choice we made again in Half 1.

If you happen to bear in mind our Preliminary Inspection, we observed that about 25% of our rows have been lacking a CustomerID. On the time, we made a strategic enterprise determination to preserve these rows. We wanted them to calculate the correct whole income and see which merchandise have been well-liked general.

For RFM evaluation, the principles change. You can not observe habits with no constant id. We are able to’t understand how “frequent” a buyer is that if we don’t know who they’re!

So, our first step in Half 3 is to isolate our “Trackable Universe” by filtering for rows the place a CustomerID exists.

Engineering the RFM Metrics

Now that we’ve got a dataset the place each row is linked to a particular particular person, we have to mixture all their particular person transactions into three abstract numbers: Recency, Frequency, and Financial.

Defining the Snapshot Date

Earlier than calculating RFM, we’d like a reference time limit, generally referred to as the snapshot date.

Right here, we take the newest transaction date within the dataset and add sooner or later. This snapshot date represents the second at which we’re evaluating buyer behaviour.

snapshot_date = df['InvoiceDate'].max() + dt.timedelta(days=1)

We added sooner or later, so prospects who purchased on the newest date nonetheless have a Recency worth of 1 day, not 0. This retains the metric intuitive and avoids edge-case issues.

Aggregating Transactions on the Buyer Degree

rfm = df.groupby(‘CustomerID’).agg({
‘InvoiceDate’: lambda x: (snapshot_date — x.max()).days,
‘InvoiceNo’: ‘nunique’,
‘Income’: ‘sum’
})

Every row in our dataset represents a single transaction. To calculate RFM, we have to collapse these transactions into one row per buyer.

We do that by grouping the information by CustomerID and making use of totally different aggregation features:

Recency: For every buyer, we discover their most up-to-date buy date and calculate what number of days have handed since then.
Frequency: We rely the variety of distinctive invoices related to every buyer. This tells us how usually they’ve made purchases.
Financial: We sum the overall income generated by every buyer throughout all transactions.

Renaming Columns for Readability

rfm.rename(columns={
'InvoiceDate': 'Recency',
'InvoiceNo': 'Frequency',
'Income': 'Financial'
}, inplace=True)py

The aggregation step retains the unique column names, which might be complicated. Renaming them makes the dataframe instantly readable and aligns it with customary RFM terminology.

Now every column clearly solutions a enterprise query:

Recency → How not too long ago did the client buy?
Frequency → How usually do they buy?
Financial → How a lot income do they generate?

Inspecting the Consequence

print(rfm.head())

The ultimate rfm dataframe comprises one row per buyer, with three intuitive metrics summarizing their habits.

Output:

Let’s stroll via this the way in which we might with NovaShop in an actual dialog.

“When was the final time this buyer purchased from us?”

That’s precisely what Recency solutions.

Take Buyer 12347:

Recency = 2
Translation: “This buyer purchased one thing simply two days in the past.”

They’re contemporary. They bear in mind the model. They’re nonetheless engaged.

Now examine that to Buyer 12346:

Recency = 326
Translation: “They haven’t purchased something in nearly a yr.”

Though this buyer spent quite a bit prior to now, they’re at present silent.

From NovaShop’s perspective: Recency tells us who’s nonetheless listening and who may want a nudge (or a wake-up name).

“Is that this a one-time purchaser or somebody who retains coming again?”

That’s the place Frequency is available in.

Look once more at Buyer 12347:

Frequency = 7
They didn’t simply purchase as soon as — they got here again time and again.

Now take a look at a number of others:

Frequency = 1
One buy, then gone.

From a enterprise perspective, frequency separates informal customers from loyal prospects.

“Who truly brings within the cash?”

That’s the Financial column.
And that is the place issues get attention-grabbing.

Buyer 12346:

Financial = £77,183.60
Frequency = 1
Recency = 326

This tells a really particular story:

A single, very massive order… a very long time in the past… and nothing since.

Now examine that to Buyer 12347:

Decrease whole spend
A number of purchases
Very latest exercise

Vital perception for NovaShop: A “high-value” buyer prior to now isn’t essentially a helpful buyer right now.

Why This View Modifications the Dialog

If NovaShop solely checked out whole income, they may focus all their consideration on prospects like 12346.

However RFM reveals us that:

Some prospects spent quite a bit as soon as and disappeared
Some spend much less however keep loyal
Some are energetic proper now and able to be engaged

This output helps NovaShop cease guessing and begin prioritizing:

Who ought to get retention emails?
Who wants reactivation campaigns?
Who’s already loyal and ought to be rewarded?

Proper now, these are nonetheless uncooked numbers.

Within the subsequent step, we’ll rank and rating these prospects, so NovaShop doesn’t should interpret rows manually. As an alternative, they’ll see clear segments like:

Champions
Loyal Clients
At-Danger
Misplaced

That’s the place this turns into an actual decision-making software — not only a dataframe.

Turning RFM Numbers Into Significant Buyer Segments

At this stage, NovaShop has a desk filled with numbers. Helpful — however not precisely decision-friendly.

A advertising workforce can’t realistically scan a whole lot or 1000’s of rows asking:

Is a Recency of 19 good or dangerous?
Is Frequency = 2 spectacular?
How a lot Financial worth is “excessive”?

Our purpose is to rank prospects relative to at least one one other and switch uncooked values into scores.

Step 1: Rating Clients by Every RFM Metric

As an alternative of treating Recency, Frequency, and Financial as absolute values, we take a look at the place every buyer stands in comparison with everybody else.

Clients with more moderen purchases ought to rating greater
Clients who purchase extra usually ought to rating greater
Clients who spend extra ought to rating greater

In apply, we do that by splitting every metric into quantiles (often 4 or 5 buckets).

Nevertheless, there’s a small real-world wrinkle. That is one thing I got here throughout whereas engaged on this challenge

In transactional datasets, it’s frequent to see:

Many purchasers with the identical Frequency (e.g. one-time patrons)
Extremely skewed Financial values
Small samples the place quantile binning can fail

To maintain issues sturdy and readable, we’ll wrap the scoring logic in a small helper operate.

def rfm_score(sequence, ascending=True, n_bins=5):
# Rank the values to make sure uniqueness
ranked = sequence.rank(methodology=’first’, ascending=ascending)

# Use pd.qcut on the ranks to assign bins
return pd.qcut(
ranked,
q=n_bins,
labels=vary(1, n_bins+1)
).astype(int)

To elucidate what’s happening right here:

We’re making a helper operate that turns a uncooked numeric column right into a clear RFM rating utilizing quantile-based binning.
First, the values are ranked. So, as a substitute of binning the uncooked values instantly, we rank them first. This step ensures distinctive ordering, even when many purchasers share the identical worth (a typical subject in RFM knowledge).
The ascending flag lets us flip the logic relying on the metric — for instance, decrease recency is healthier, whereas greater frequency and financial values are higher.
Subsequent, we’re making use of quantile-based binning. qcut splits the ranked values into n_bins equally sized teams. Every buyer is assigned a rating from 1 to five (by default), the place the rating represents their relative place inside the distribution.
Lastly, the outcomes shall be transformed to integers for simple use in evaluation and segmentation.

In brief, this operate offers a sturdy and reusable manner to attain RFM metrics with out operating into duplicate bin edge errors — and with out overcomplicating the logic.

Step 2: Making use of the Scores

Now we are able to rating every metric cleanly and constantly:

# Assign R, F, M scores
rfm['R_Score'] = rfm_score(rfm['Recency'], ascending=False) # Latest purchases = excessive rating
rfm['F_Score'] = rfm_score(rfm['Frequency']) # Extra frequent = excessive rating
rfm['M_Score'] = rfm_score(rfm['Monetary']) # Larger spend = excessive rating

The one particular case right here is Recency:

Decrease values imply more moderen exercise
So we reverse the rating with ascending=False
The whole lot else follows the pure “greater is healthier” rule.

What This Means for NovaShop

As an alternative of seeing this:

Recency = 326
Frequency = 1
Financial = 77,183.60

NovaShop now sees one thing like:

R = 1, F = 1, M = 5

That’s immediately extra interpretable:

Not latest
Not frequent
Excessive spender (traditionally)

Step 3: Making a Mixed RFM Rating

Now we mix these three scores right into a single RFM code:

rfm['RFM_Score'] = (
rfm['R_Score'].astype(str) +
rfm['F_Score'].astype(str) +
rfm['M_Score'].astype(str)
)

This produces values like:

555 → Greatest prospects
155 → Excessive spenders who haven’t returned
111 → Clients who’re doubtless gone

Every buyer now carries a compact behavioral fingerprint. And we’re not executed but.

Translating RFM Scores Into Buyer Segments

Uncooked scores are good, however let’s be sincere: no advertising supervisor desires to have a look at 555, 154, or 311 all day.

NovaShop wants labels that make sense at a look. That’s the place RFM segments are available.

Step 1: Defining Segments

Utilizing RFM scores, we are able to classify prospects into significant classes. Right here’s a typical strategy:

Champions: Prime Recency, high Frequency, high Financial (555) — your greatest prospects
Loyal Clients: Common patrons, is probably not spending probably the most, however preserve coming again
Huge Spenders: Excessive Financial, however not essentially latest or frequent
At-Danger: Used to purchase, however haven’t returned not too long ago
Misplaced: Low scores in all three metrics — doubtless disengaged
Promising / New: Latest prospects with decrease frequency or financial spend

This transforms summary numbers right into a narrative that advertising and administration can act on.

Step 2: Mapping Scores to Segments

Right here’s an instance utilizing easy conditional logic:

def rfm_segment(row):
if row['R_Score'] >= 4 and row['F_Score'] >= 4 and row['M_Score'] >= 4:
return 'Champions'
elif row['F_Score'] >= 4:
return 'Loyal Clients'
elif row['M_Score'] >= 4:
return 'Huge Spenders'
elif row['R_Score'] <= 2:
return 'At-Danger'
else:
return 'Others'
rfm['Segment'] = rfm.apply(rfm_segment, axis=1)

Now every buyer has a human-readable label, making it instantly actionable.

Let’s evaluation our outcomes utilizing rfm.head()

Step 3: Turning Segments into Technique

With labeled segments, NovaShop can:

Reward Champions → Unique offers, loyalty factors
Re-engage Huge Spenders & At-Danger prospects → Personalised emails or reductions
Focus advertising correctly → Don’t waste effort on prospects who’re really misplaced

That is the second the place knowledge turns into technique.

What NovaShop Ought to Do Subsequent (Key Takeaways & Suggestions)

At the beginning of this evaluation, NovaShop had a well-recognized drawback:
A number of transactional knowledge, however restricted readability on buyer behaviour.

By making use of the RFM framework, we’ve turned uncooked buy historical past into a transparent, structured view of who NovaShop’s prospects are — and the way they behave.

Now let’s speak about what to truly do with it.

1. Defend and Reward Your Greatest Clients

Champions and Loyal Clients are already doing what each enterprise desires:

They purchase not too long ago
They purchase usually
They generate constant income

These prospects don’t want heavy reductions — they want recognition.

Beneficial actions:

Early entry to gross sales
Loyalty factors or VIP tiers
Personalised thank-you emails

The purpose right here isn’t acquisition, it’s retention.

2. Re-Interact Excessive-Worth Clients Earlier than They’re Misplaced

Probably the most harmful phase for NovaShop isn’t “Misplaced” prospects.
It’s At-Danger and Huge Spenders.

These prospects:

Have proven clear worth prior to now
However haven’t bought not too long ago
Are one step away from churning utterly

Beneficial actions:

Focused win-back campaigns
Personalised provides (not blanket reductions)
Reminder emails tied to previous buy habits

Successful again an present buyer is nearly at all times cheaper than buying a brand new one.

3. Don’t Over-Put money into Really Misplaced Clients

Some prospects will inevitably churn. RFM helps NovaShop determine these prospects early and keep away from spending advert finances, reductions and advertising effort on customers who’re unlikely to return. This isn’t about being chilly — it’s about being environment friendly.

4. Use RFM as a Dwelling Framework, Not a One-Off Evaluation

The true energy of RFM comes when it’s:

Recomputed month-to-month or quarterly
Built-in into dashboards
Used to trace motion between segments over time

For NovaShop, this implies asking questions like:

What number of At-Danger prospects grew to become Loyal this month?
Are Champions rising or shrinking?
Which campaigns truly transfer prospects up the ladder?

RFM turns buyer behaviour into one thing measurable and trackable.

Closing Ideas: Closing the EDA in Public Sequence

Once I began this EDA in Public sequence, I wasn’t making an attempt to construct the right evaluation or show superior methods. I wished to decelerate and share how I truly suppose when working with actual knowledge. Not the polished model, however the messy, iterative course of that often stays hidden.

This challenge started with a loud CSV and plenty of open questions. Alongside the way in which, there have been small points that solely surfaced as soon as I paid nearer consideration — dates saved as strings, assumptions that didn’t fairly maintain up, metrics that wanted context earlier than they made sense. Working via these moments in public was uncomfortable at instances, but additionally genuinely helpful. Every correction made the evaluation stronger and extra sincere.

One factor this course of strengthened for me is that the majority significant insights don’t come from complexity. They arrive from slowing down, structuring the information correctly, and asking higher questions. By the point I reached the RFM evaluation, the worth wasn’t within the formulation themselves — it was in what they compelled me to confront. A buyer who spent quite a bit as soon as isn’t essentially helpful right now. Recency issues. Frequency issues. And none of those metrics imply a lot in isolation.

Ending the sequence with RFM felt deliberate. It sits on the level the place technical work meets enterprise considering, the place tables flip into conversations and numbers flip into selections. It’s additionally the place exploratory evaluation stops being purely descriptive and begins changing into sensible. At that stage, the purpose is now not simply to know the information, however to determine what to do subsequent.

Doing this work in public modified how I strategy evaluation. Writing issues out compelled me to clarify my reasoning, query my assumptions, and be snug exhibiting imperfect work. It jogged my memory that EDA isn’t a guidelines you rush via — it’s a dialogue with the information. Sharing that dialogue makes you extra considerate and extra accountable.

This can be the ultimate a part of the EDA in Public sequence, but it surely doesn’t really feel like an endpoint. The whole lot right here might evolve into dashboards, automated pipelines, or deeper buyer evaluation.

And in the event you’re a founder, analyst, or workforce working with buyer or gross sales knowledge and making an attempt to make sense of it, this type of exploratory work is commonly the place the most important readability comes from. These are precisely the sorts of issues I take pleasure in working via — slowly, thoughtfully, and with the enterprise context in thoughts.

If you happen to’re documenting your individual analyses, I’d like to see the way you strategy it. And in the event you’re wrestling with related questions in your knowledge and need to speak via them, be at liberty to achieve out on any of the platforms under. Good knowledge conversations often begin there.

Thanks for following alongside!

Medium

Twitter

YouTube