Stop Chasing Tools, Start Using Frameworks — From CRISP-DM to DMAIC: lessons from a banking data professional

5 min readJust now

–

For a long time, I thought “being good with data” meant stacking more tools.

More SQL. More Python. More dashboards. If I could engineer cleaner pipelines and prettier charts, I’d be safe. Then I spent a few years in banking and finance.

I watched “perfect” models die in review meetings. I saw beautiful dashboards get ignored. I saw scrappy, half-manual processes survive for years because they fit how people actually made decisions.

Slowly, it clicked:

The real leverage in analytics isn’t the toolset. It’s how you think about the problem.

That shift didn’t come from another course or another library. It came from a handful of old, almost boring frameworks that quietly rewired the way I approach data work.

This isn’t a textbook walk-through. It’s how seven frameworks changed how I think.

Press enter or click to view image in full size

Created By Dewank Mahajan — 7 Analytics Frameworks That Changed How I Think About Data Problems

CRISP-DM: The Day I Got Embarrassed

CRISP-DM was my first real slap in the face.

I’d been obsessing over the “Modeling” box: accuracy, ROC curves, clever features. CRISP-DM begins two steps earlier: Business Understanding and Data Understanding.

In banking, that difference is everything.

“Increase card spend” is not one problem; it’s three different problems depending on whether you sit in marketing, risk, or compliance. The same model can be genius in one context and reckless in another.

CRISP-DM forced me to ask, before I opened a notebook:

• What decision will this actually influence?
• Who owns that decision?
• How will success be measured in the real world, not just on a validation set?

Once I respected those questions, a lot of model drama disappeared. I built fewer “smart” solutions and more useful ones.

SEMMA: Learning to Love the Sample

SEMMA — Sample, Explore, Modify, Model, Assess — snuck into my life via SAS veterans.

What stuck with me wasn’t the full acronym. It was that very first word: Sample.

In regulated environments, the instinct is to pull everything: all years, all products, all customers. It feels rigorous. It’s usually paralysing.

SEMMA gave me permission to start smaller. Take a clean, representative slice of the portfolio. Explore behavior. Try transformations. Test a rough model. Learn something. Then scale.

It sounds trivial, but it changed my default from “build a data universe” to “get a fast, honest signal.” Progress over perfection.

OODA Loop: When Analytics Joined the Fight

The OODA loop — Observe, Orient, Decide, Act — comes from military strategy, not data science.

I only started caring about it when we were tracking shifts in a loan book as macro conditions moved. Suddenly, analytics wasn’t a static report. It was part of a live feedback loop with risk, product, and treasury.

OODA gave me a language for that:

First we observe the signals: delinquency, utilization, churn, external indicators. Then we orient: what’s noise, what’s structural, what’s seasonal? We decide: raise limits, tighten criteria, change collections strategy — or wait. And then we act.

And then we loop.

Dashboards stopped being “deliverables” and became instruments. Models stopped being projects and became sensors. The question changed from “What can we analyze?” to “How quickly can we close this loop between what we see and what we do?”

KDD: Respecting the Messy Middle

Knowledge Discovery in Databases (KDD) is the first framework that made me fall in love with the middle of the process.

We glamorize the start (strategy, questions) and the end (insights, impact). KDD quietly insists that the unsexy stages — selection, preprocessing, transformation — are where the fate of a project is usually decided.

In banking, that’s painfully true.

Pick the wrong population and your credit model is biased before it trains. Ignore messy product hierarchies and your profitability analysis becomes a guessing game. Treat preprocessing as an afterthought and all downstream sophistication is built on sand.

KDD made me less apologetic about spending serious time there. The middle isn’t a hurdle on the way to “real data science.” It is real data science.

PDCA: The Rhythm of Boring Excellence

Plan, Do, Check, Act (PDCA) sounded like corporate wallpaper the first time I heard it.

Then I inherited a recurring monthly risk report.

Everything changed, all the time: data quality, definitions, thresholds, business priorities. Fixes happened in chat threads and late-night patches. There was no memory, just reactions.

PDCA gave me a simple rhythm:

Plan what we’ll change this cycle and why. Do the change in code or process. Check what actually happened. Act by locking in the improvement or rolling it back.

No heroics. No big-bang redesigns. Just small, deliberate cycles. Over time, that rhythm turned a fragile, stressful month-end into something almost boring.

In analytics, “boring” is often what success looks like.

TDSP: Analytics as a Team Sport

The Team Data Science Process (TDSP) punched a small hole in my ego.

For years I enjoyed being the “end-to-end” person. I could meet stakeholders, write SQL, build models, and ship dashboards. It felt efficient — and a little heroic.

TDSP doesn’t argue against that skillset. It just reminds you that serious analytics lives inside teams, not CVs.

Standard project structures, shared artifacts, consistent folder layouts, version control, defined roles: it all feels bureaucratic until you’re in your third audit, or your lead data scientist leaves, or your model misbehaves in production and nobody knows why.

TDSP nudged me from “look what I built” to “how do we build this so others can own, debug, and extend it a year from now?” Less glamorous. Much healthier.

DMAIC: Putting Handles on Messy Problems

DMAIC — Define, Measure, Analyze, Improve, Control — arrived in my life on a Six Sigma slide I mostly ignored.

It earned my respect during a simple but painful situation: a key metric showed three different values in three different reports across the bank. Everyone was convinced they were right. No one trusted anyone else’s numbers.

We could have argued for weeks.

DMAIC gave us handles.

We defined the problem clearly. We measured how often and how badly it showed up. We analyzed the causes — different filters, timings, data sources. We agreed improvements: a single logic, one source of truth. Finally, we put controls in place so it didn’t quietly fragment again.

What I like about DMAIC is that it doesn’t treat data issues as background noise. It treats them as improvement projects with a beginning, a middle, and an end.

The Real Shift

I don’t follow any of these frameworks religiously. Real projects are messy. We loop back, skip steps, stitch two frameworks together in one sprint.

But they changed my inner dialogue.

CRISP-DM makes me ask “What business problem is this really?” SEMMA whispers “Start small, learn fast.” OODA reminds me “This analysis lives in a loop, not a slide deck.” KDD says “Don’t rush the middle; that’s where truth leaks in or out.” PDCA keeps me improving the process, not just the product. TDSP pushes me to build things teams can live with. DMAIC turns vague frustration into structured action.

Tools come and go. Frameworks shape how you see.

___________________________

🚀 Let’s Connect

👉 Follow me here on Medium for more ML insights & case studies. 💡 Join me on LinkedIn for professional takes. 🐦 X • 🎥 TikTok • 📸 Instagram → daily drops & quick hits.