Project Planning for Data Scientists

Why Most Data Science Projects Fail

Industry surveys have repeatedly found that a large share of data science and ML projects never make it into production. The usual reasons are rarely technical. They are:

• The problem was vague, so nobody could agree whether the result was useful
• The data needed did not exist, was not accessible, or was far messier than expected
• The model answered a question nobody was going to act on
• The project took so long that the business moved on

Good planning prevents all four. Before you open a notebook, spend time scoping the project. This lesson gives you a framework you can use in your portfolio projects and at work, whether at a Lagos fintech or a global company.

Step 1: Start With the Decision, Not the Data

Ask the stakeholder: "What decision will you make differently with this result?"

If nobody can name a decision, the project is not ready.

Step 2: Define Success in Numbers

Agree on two kinds of metrics before you start:

• Business metric: what the organisation cares about, such as reducing churn from 4.2% to 3.5% a month, or cutting fraud losses by N50 million a quarter
• Model metric: how you will judge the model, such as recall of at least 70% at a precision of at least 40%

Also define the baseline: what happens today without your model? A model is only valuable if it beats the current process, which might be a simple rule or a manager's judgement.

Step 3: Audit the Data

For every data source, answer:

Data leakage is one of the most common mistakes: using information that would not be known at the moment of prediction. Using "number of calls to the cancellation line" to predict churn looks brilliant in testing and is useless in production, because by then the customer has already decided to leave.

Step 4: Scope the Deliverables

Be explicit about what you will hand over:

• An exploratory analysis with findings (days to weeks)
• A dashboard that updates regularly
• A model plus documentation
• A deployed API or batch scoring job integrated with a product
• A presentation with recommendations

Start with the smallest version that proves value, a minimum viable analysis. A simple model with a CSV list of high-risk customers that the retention team can use next week beats a perfect deployed system six months from now.

Step 5: Plan the Timeline and Risks

A typical four-week portfolio or first-iteration project:

List the top risks (data arrives late, labels are unreliable, stakeholder availability) and a mitigation for each.

Step 6: Write a One-Page Project Brief

Put it all on one page and get the stakeholder to agree before you build:

PROJECT: Reduce prepaid churn in Lagos
DECISION: Which subscribers receive a retention bundle each week
SUCCESS: Monthly churn in the target group falls from 4.2% to 3.5%
MODEL METRIC: Recall >= 70% at precision >= 40% on a held-out month
BASELINE: Current rule - offer to anyone inactive for 14+ days
DATA: 12 months of usage, recharge and complaint logs (access: approved by the CRM lead)
DELIVERABLE (v1): Weekly CSV of the top 5,000 at-risk subscribers + a slide deck
TIMELINE: 4 weeks
RISKS: Complaint data only goes back 6 months -> test the model with and without it

This brief is also a great portfolio artefact. Recruiters love seeing that you think about the business, not only the algorithm.

Try it: Use the planner below to scope a project of your own. It flags common gaps before you start.