Data tells you the what.
I explain the why behind. Data tell the what.
I explain the why.

Data Scientist · Lagos, Nigeria
Open to full-time roles and contracts

See my work About me

Case Study

Quantifying Marketing ROI with Synthetic ControlMarketing ROI with Synthetic Control

Using Synthetic Control to isolate and measure the true impact of a marketing campaign in Port Harcourt.

Case Study

Free delivery and its true impact on spendingFree delivery and its impact on spending

Using Regression Discontinuity to prove that free delivery actually triggers customers to spend 24% more per order.

Case Study

Separating campaign impact from coincidence

Applying Difference-in-Differences to marketing data to determine what truly drove user acquisition.

Dashboard

Baraa's YouTube channel dashboard

A live analytics dashboard synced to Google Sheets. Tracks views, engagement, and retention for Baraa's channel in real time.

Product

A free CV tool that actually worksA CV Tool That Actually Works

I built a custom web app to automate my personal CV review process, helping job seekers bypass ATS filters.

The Origin

Paying attention.

I grew up watching my mother run her shop. She knew which customers would come back before they did. She never had a spreadsheet. She just paid attention.

That stuck with me. The idea that behavior tells you everything, if you know how to read it. I went on to study Mathematics at Federal University of Technology Owerri, graduated, and eventually found my way into data.

The Science

Beyond correlation.

I analyze data, build visual stories around it, and turn those stories into insights that help products grow. My work sits at the intersection of product analytics and behavioral economics.

I have been studying applied econometrics and causal methods seriously, using techniques like Regression Discontinuity and Difference-in-Differences to answer questions that experiments cannot.

Time with Others

Betting Apps in NigeriaBetting Apps

Work and Identity

Boko Haram Story

App

OneGoodCV Builder OneGoodCV Builder

I was tired of writing CVs for everyone so I built a free tool that helps anyone write their own. Clean, simple, done.

Dashboard

Baraa YouTube Dashboard Baraa's Dashboard

A live analytics dashboard synced to Google Sheets. Tracks views, engagement, and retention for Baraa's channel in real time.

Course

Lazy Data Engineering Lazy Data Engineering

How I collected Baraa's YouTube data using the YouTube API. A practical walkthrough for people who want to build real data pipelines without overcomplicating it.

Got it. I will get back to you soon.

404

This page does not exist.

The page you are looking for may have moved, been renamed, or never existed. The data on this one is clear.

Go home Go back

Causal Inference

Quantifying Marketing ROI with Synthetic ControlMarketing ROI with Synthetic Control

The Problem

The food delivery company ChowChow launched a marketing campaign in Port Harcourt to increase daily orders. The company spent a significant amount of money and now needs to know if that investment was successful. To find the answer, we must compare Port Harcourt to other cities that did not receive the campaign. However, Lagos and Abuja are much larger and grow at different speeds. This makes it very difficult to tell if a change in orders was caused by our marketing or if it was just a natural trend in the market.

The Big Question

How many extra orders did the marketing campaign actually produce each day in Port Harcourt?

View the Python notebook

The Method

A common way to solve this is a naive approach where we simply look at the average number of orders before the campaign and compare it to the average after the campaign. This method usually fails because it ignores what is happening in the rest of the country. For example, if orders go up in every city because of a national holiday, the naive approach would incorrectly credit our marketing for that growth. We need a method that looks at other cities to act as a benchmark so we can see what would have happened if we never ran the campaign.

Daily Order Trends Across Major Markets Daily Order Trends
Across Major Markets

Total orders per day (in thousands)Total orders per day, Jan 2023 - Jul 2025

Why Difference-in-Differences (DiD) Failed Why DiD Failed

Difference-in-Differences failed in our case because it relies on the Parallel Trends Assumption. This means the control group and Port Harcourt must move together perfectly before the campaign starts. As we saw in the EDA chart, this is not true. Lagos is in a world of its own. It is much larger and grows at a different speed. Because the control group was moving faster than Port Harcourt before the campaign even began, DiD created a biased benchmark. It expected Port Harcourt to keep up with the fast growth of Lagos, which made our actual success look like a failure.

Our Solution: Synthetic Control

Since DiD failed because the average city was a bad mirror, we used Synthetic Control to build a custom one. We used Synthetic Control to build a Digital Twin of Port Harcourt using other cities as Donors. Instead of using a simple average, this method allows us to pick and choose exactly which cities are most like Port Harcourt to create a more realistic comparison.

Analysis Setup Summary

Setting	Configuration
Treated City	Port Harcourt
Donor Cities	Lagos, Abuja, Ibadan, Kano
Campaign Start Day	1st December 2024 (Day 700)
Pre-campaign Duration	700 days
Post-campaign Duration	300 days
Total Analysis Period	1000 days

Optimiser Weights

Donor City	Weight Contribution
Lagos	0.1754 (17.5%)
Abuja	0.1479 (14.8%)
Ibadan	0.3658 (36.6%)
Kano	0.3109 (31.1%)

Pre-Campaign Fit Check

Metric	Value
PH average daily orders before campaign	316.7
Pre-campaign RMSPE	13.88 orders/day
RMSPE as percentage of baseline	4.4%

Full Period Results

Metric	Value
PH average daily orders after campaign	438.2
Synthetic PH average after campaign	389.9
Estimated daily orders from campaign	48 orders/day

Port Harcourt did better because of the campaign Port Harcourt did better
because of the campaign

Daily orders, actual vs. synthetic; RMSPE 4.4%

Is This a Fluke?

We performed a check called a Placebo Test to prove that our result of 48 orders was real and not just luck. We pretended that the other cities received the campaign instead of Port Harcourt and ran the same model on them. Since those cities did not actually have a campaign, their results should be close to zero.

Placebo test results

Pre-period fit is how closely each city tracks its synthetic twin before the campaign cutoff. Post-campaign lift is the gap after the fake intervention. The row we care about is the real campaign city.

City	Pre-period fit	Post-campaign lift
Port Harcourt	4.4%	48 orders/day
Lagos	6.7%	-163 orders/day
Abuja	4.9%	-22 orders/day
Ibadan	4.1%	-4 orders/day
Kano	8.9%	2 orders/day

Port Harcourt

Pre-period fit 4.4%

Post-campaign lift 48 orders/day
Lagos

Pre-period fit 6.7%

Post-campaign lift -163 orders/day
Abuja

Pre-period fit 4.9%

Post-campaign lift -22 orders/day
Ibadan

Pre-period fit 4.1%

Post-campaign lift -4 orders/day
Kano

Pre-period fit 8.9%

Post-campaign lift 2 orders/day

Highlighted row: actual campaign city (not a placebo).

Our test showed that Port Harcourt was the only city with a significant positive jump. While Lagos showed a result of -163 orders, it is important to remember that Lagos is a giant. When we look at the percentages, the difference is clear:

Port Harcourt (The Signal): A jump of 48 on a baseline of 317 is a 15% increase. This is a clear sign that the campaign worked.
Lagos (The Noise): A drop of 163 on a baseline of 1,200 is only a 13% error.

In a giant market like Lagos, a 13% fluctuation is normal "noise" or a side effect of the model trying to build a twin for a city that has no equal. Most importantly, Port Harcourt was the only city to break upward. The fact that the jump in Port Harcourt was so large that it even caused the models for other cities to over-predict their growth is the ultimate proof that the lift was real and unique to Port Harcourt.

Summary

The marketing campaign in Port Harcourt was a clear success. While standard methods like Difference-in-Differences failed due to the size and speed of Lagos, our Synthetic Control model provided a reliable answer. By creating a Digital Twin that was 95.6% accurate, we proved that the campaign generated 48 additional orders every day. This analysis gives ChowChow a clear and data-driven way to measure the value of their marketing investments.

To build the Digital Twin, we needed to answer one key question: how much each donor city should contribute. If Lagos is twice as similar to Port Harcourt as Kano, then Lagos should have more influence. These influence values are called weights. We used a mathematical optimizer to find the best weights for each city. Here is how it worked:

1. The Loss Function (Mean Squared Error)

We gave the model a target called a Loss Function. In our project, we used Mean Squared Error (MSE). The job of the loss function is to calculate the "penalty" for being wrong.

It takes the difference between the Twin and the Real city for every single day.
It squares that difference so that even small errors are punished.
The optimizer's only goal is to change the weights until this "penalty" score is as close to zero as possible.

2. The Scaling Secret (Mean-Normalization)

This is the most important part of our technical approach. Without scaling, the optimizer is "blinded" by the size of Lagos. We used a method called Mean-Normalization.

The Math:
We calculate the average (Mean) of each city for the entire 700-day period. Then, we divide every daily order count by that average.

What the Scaling Does (An Example):
Imagine we have two donor cities: Lagos (Huge) and Kano (Small). On a specific day, both cities grow by 10%.

City	Average	Orders	Calculation	Scaled
Lagos	1,200	1,320	1,320 ÷ 1,200	1.10
Kano	10	11	11 ÷ 10	1.10

Why this matters:
Before scaling, a 10% error in Lagos was 120 orders, while a 10% error in Kano was only 1 order. The optimizer would ignore Kano's trend entirely because its numbers were too small to affect the total "penalty" score. By scaling, both cities now look like they have a value of 1.10. This gives Kano an equal "voice" in the model. It forces the optimizer to look at the Trend (the wiggles in the line) rather than the Volume (the size of the city).

3. The Constraints

We set strict rules so the model stayed realistic:

Sum to 1.0: The total donor weights must equal 100%. We cannot create a twin that is "larger" than a full city.
No Negatives: A city cannot have a negative weight. We only want to add trends together, not subtract them.

Following these rules, the optimizer gave us our final weights: Lagos: 0.1754, Abuja: 0.1479, Ibadan: 0.3658, and Kano: 0.3109. These perfectly equal 1.0.

	Model Features	CoefficientEstimated effect	P > \|t\|p-value
1	const	4,299	0.000
2	treatment	4.33	0.977
3	post	783	0.000
4	treat_post	1,450	0.000
5	urban	1,837	0.000
6	promotion_flag	-69.7	0.534

Data tells you the what.I explain the why behind. Data tell the what.I explain the why.

Quantifying Marketing ROI with Synthetic ControlMarketing ROI with Synthetic Control

Free delivery and its true impact on spendingFree delivery and its impact on spending

Separating campaign impact from coincidence

Baraa's YouTube channel dashboard

A free CV tool that actually worksA CV Tool That Actually Works

Translating data into clear insights. Translating datainto clear insights. Translating data into clear insights.

Paying attention.

Beyond correlation.

Investigating what actually shapes us. Investigating whatactually shapes us. Investigating what actually shapes us.

Time with Others

Betting Apps in NigeriaBetting Apps

Work and Identity

Boko Haram Story

Things I built because they needed to exist. Things I built because they needed to exist. Things I built because they needed to exist.

Tell me what you are trying to solve. Tell me what you are trying to solve. Tell me what you are trying to solve.

This page does not exist.

Quantifying Marketing ROI with Synthetic ControlMarketing ROI with Synthetic Control

The Problem

The Big Question

The Method

Why Difference-in-Differences (DiD) Failed Why DiD Failed

Our Solution: Synthetic Control

Analysis Setup Summary

Optimiser Weights

Pre-Campaign Fit Check

Full Period Results

Is This a Fluke?

Placebo test results

Summary

Behind the Scene: How We Got the WeightsHow We Got the Weights

1. The Loss Function (Mean Squared Error)

2. The Scaling Secret (Mean-Normalization)

3. The Constraints

Free delivery and its true impact on spendingFree delivery and its impact on spending

The Problem

The Big Question

The Method

Model Results

Business Takeaway

Separating campaign impact from coincidence

The Problem

Simple Methods Do Not Work

The Method

Assumptions

Regression Results

The Result

A Free CV Tool That Actually WorksA CV Tool That Actually Works

The Problem I Kept Seeing

Testing the Waters

The Solution: OneGoodCV

What I Built

OneGoodCV in Action

Building It (The Hard Part)

What those 30 days looked like

My Research Process

Key insights I discovered

What Made It Different

Early Results

Key Lessons Learned

What Is Next

Skills This Project Showcased

Data Collection

Getting Your API Key

Watch How It Works

Your Target YouTube Channel

How to Find the Channel ID

Watch the Demo

Collecting YouTube Data

Watch How It Works

Step by Step Setup

Step 1: Open the Script Editor

Step 2: Add the Script

Step 3: Set Up the Daily Trigger

Run It for the First Time

Data tells you the what.
I explain the why behind. Data tell the what.
I explain the why.

Translating data
into clear insights. Translating data
into clear insights. Translating
data into clear
insights.

Investigating what
actually shapes us. Investigating what
actually shapes us. Investigating
what actually
shapes us.

Things I built because
they needed to exist. Things I built because
they needed to exist. Things I built
because they
needed to exist.

Tell me what you
are trying to solve. Tell me what you
are trying to solve. Tell me what
you are trying
to solve.