Six months ago, I was in a meeting at Morningstar where someone casually mentioned our fund recommendation algorithm uses "machine learning." I nodded like I understood. I didn't.
I went home, watched a YouTube video titled "Machine Learning Explained in 5 Minutes," felt confident for exactly 47 seconds, then realized I'd just learned that ML recognizes cats in photos. That wasn't helpful. That wasn't even close to what we were actually doing.
So I did what any data analyst who's tired of pretending should do — I asked stupid questions. For three months. And what I learned completely changed how I think about my job, my data, and honestly, why companies like Morningstar, Zerodha, and even your CRED app work the way they do.
Here's what I got wrong first. Then what actually stuck.
What I Thought Machine Learning Was (And Why I Was Wrong)
My first mental model of ML was basically: upload data, press button, AI does magic, profit.
I thought it was one thing. A monolith. A technology that either worked perfectly or didn't work at all. I imagined algorithms as almost sentient — like they "learned" the way I learn something from a textbook, retained it, and could apply it to new situations flawlessly.
The "Magic Box" Misconception
This was the biggest trap. I genuinely believed that machine learning was this black-box technology where you feed in messy data and out comes perfect predictions. Netflix knows what you'll watch. Google knows what you'll search. Surely the algorithm just... knows things.
What I didn't realize? Those systems fail constantly. Netflix recommends trash sometimes. Google search results are sometimes useless. The algorithm isn't magic — it's just statistically better than random guessing. And that's actually the entire point.
When I started analyzing mutual fund performance at Morningstar, I kept waiting for our ML model to nail predictions consistently. It didn't. Some quarters it got 62% accuracy. Some months, 58%. I was frustrated. My manager said, "That's actually great." That confused me for weeks.
The "Robots Are Learning Like Humans" Myth
I also thought "machine learning" meant machines were actually learning in any meaningful sense. Like, the algorithm was thinking, understanding, maybe even reasoning.
Nope.
The word "learning" is maybe the worst marketing decision in tech history. Because machines aren't learning. They're pattern-matching at scale. Finding correlations. It's statistics dressed up in a blazer and told to go to the office. Nothing more.
What Machine Learning Actually Is (The Honest Version)
Okay, so here's what I actually figured out.
Machine learning is a way to find patterns in data without explicitly programming every rule. That's it. That's the whole thing.
Normally, if you wanted a computer to recognize whether a mutual fund will outperform in the next quarter, you'd manually write rules: "If expense ratio is below X, AND Sharpe ratio is above Y, AND the manager has been around for Z years, THEN predict outperformance." Every single rule, you code it in.
But what if the real pattern is way more complex? What if it's actually: "This specific combination of expense ratio, manager tenure, and market volatility, blended in a non-obvious way, predicts performance"? You'd need 10,000 hand-coded rules. Impossible.
That's where ML comes in. You feed the system historical data — thousands of examples of funds and whether they actually outperformed. The algorithm finds the patterns you couldn't see. It builds a model (basically a giant mathematical equation) that captures those relationships.
Then when you feed it a new fund, it runs that fund's data through the equation and gives you a prediction.
Here's the part nobody tells you: the algorithm doesn't understand anything. It has no idea what "mutual fund" means. It's just finding mathematical correlations between numbers. If you renamed "expense ratio" to "pizza consumption," the algorithm wouldn't care. It works either way.
How It Actually Works (Without the BS)
Step 1: Collect Messy, Historical Data
You gather examples. A Zerodha trading bot needs thousands of stock price movements with actual trading outcomes. A CRED credit scoring model needs thousands of loan applicants with their characteristics and whether they eventually defaulted.
The quality and size of this data? Absolutely critical. Garbage in, garbage out. If your historical data is biased, your model will be biased.
Step 2: Split the Data and Pick the Algorithm
You divide your data: 80% for training, 20% for testing. (Sometimes other ratios, depending on what you're doing.)
Then you pick an algorithm. Common ones for data like ours: decision trees, neural networks, random forests, linear regression. Each one finds patterns differently. A neural network is like a super-complex web of interconnected nodes. A random forest builds hundreds of decision trees and averages their votes.
Here's the thing: nobody just knows which algorithm will work best. You have to try multiple, test them, compare. This is not elegant. This is trial and error with statistics.
Step 3: Train the Model (The Actual "Learning" Part)
You feed the 80% training data into the algorithm. It looks for patterns. It builds a mathematical model. This process adjusts weights and parameters thousands of times to minimize error — basically, to get the predictions as close to the actual historical outcomes as possible.
The algorithm doesn't learn in any conscious sense. It's literally just doing calculus. Optimization. Finding the values that minimize the difference between "what it predicted" and "what actually happened."
Step 4: Test It on New Data
Now you run your model on the 20% test data it's never seen. How well does it predict? If it's 62% accurate, great — that's probably better than a coin flip. If it's 51% accurate, terrible. If it's 99% accurate, suspicious. (You might have overfit.)
Step 5: Deploy and Monitor
You put the model into production. It makes predictions on real data. And here's what everyone skips over: you keep watching it. Because the patterns that worked in historical data don't always hold up in the real world. Markets change. Human behavior changes. Your model degrades.
At Morningstar, we monitor our fund recommendation model every single week. If accuracy drops below a threshold, we either retrain it with newer data or go back to the drawing board. This isn't a one-time thing. It's ongoing maintenance.
Real Examples You'll Recognize
Let me give you examples from actual Indian tech products you use.
Groww's fund recommendations: You open the app, they suggest index funds or mutual funds. That could be rule-based (if you're a 25-year-old salaried professional, we recommend equity funds). But Groww probably also uses ML — looking at thousands of users like you and which funds they bought, which ones they held, which ones performed well for them. Finding patterns. Recommending accordingly.
PhonePe's fraud detection: When you try to send ₹50,000 to someone at 2 AM for the first time, PhonePe blocks it. Why? Rules, partly. But also ML. The system has learned patterns from millions of transactions about what looks fraudulent. Your transaction matches the pattern.
Zerodha's market data and tools: When you see technical indicators calculated instantly on their charts, or when they flag unusual volume spikes, that's not magic. That's algorithms finding patterns in price and volume data from decades of stock movements.
CRED's credit scoring: You apply for a loan. The traditional bank uses your credit score. CRED looks at your bill payment history, your spending patterns (from their payment data), your frequency of paycheck deposits. Patterns. Correlations. Probability that you'll repay.
None of these are thinking or understanding. They're all finding correlations at speed and scale.
The Limits Nobody Talks About
This is important. Machine learning is incredible at finding patterns in historical data and making probabilistic predictions about the future.
It is absolutely terrible at:
Understanding causation. ML finds correlation. It doesn't know why. If an ML model predicts that mutual fund investors who drink coffee are more likely to hold funds long-term, it doesn't mean coffee causes fund loyalty. Maybe coffee drinkers are older, on average. Or maybe they're more disciplined. But the algorithm has no idea. It just knows the correlation exists.
Handling completely new situations. If your ML model learned from 2015-2023 data, it might fail spectacularly in 2024 if something unprecedented happens. COVID tanked markets in ways no historical pattern could predict. The models struggled.
Explaining itself clearly. Deep neural networks especially are black boxes. They make good predictions, but ask them *why* — why did you predict this outcome? — and they can't answer. "The weights came out this way." That's not an explanation.
Handling sparse or imbalanced data. If your training data has 99% of people who didn't default on loans and 1% who did, your model will be terrible at predicting defaults. It learns that "always predict no default" is 99% accurate. Technically true. Useless.
| What ML Does Well | What ML Struggles With |
|---|---|
| Finding patterns in historical data | Understanding *why* the pattern exists |
| Making probabilistic predictions | Predicting unprecedented events |
| Processing data at massive scale | Working with small or imbalanced datasets |
| Improving with more data | Making transparent, explainable decisions |
| Finding non-obvious correlations | Knowing when it's wrong (before real consequences) |
My Perspective After Six Months
Here's what surprised me: machine learning isn't revolutionary in the way I thought. It's not because the algorithms are magic. It's revolutionary because we can now process data at scale and find patterns faster than humans ever could.
At Morningstar, I realized something that changed how I work. Our best fund recommendations don't come from a single sophisticated neural network. They come from combining multiple simple models — linear regressions, decision trees, human expert judgment — and trusting that ensemble approach more than trusting any single algorithm. The best ML work isn't fancy. It's humble.
I also got this wrong: I thought implementing ML meant the hard part was over. Actually, implementation is maybe 10% of the work. The hard parts are: getting clean data (months), picking the right metric (weeks of debate), monitoring drift (forever), and honestly? Explaining to stakeholders that your 65% accurate model is actually valuable, even though it seems like it should be much better.
What I'd do differently now? I'd spend less time worrying about algorithm choice and way more time on data quality and problem definition. I've seen fancy algorithms fail on messy data, and simple algorithms succeed on clean, well-defined problems. The quality of your inputs and clarity about what you're trying to predict matter infinitely more than whether you use a random forest or a neural network.
Final Thoughts
Machine learning isn't magic. It's not AGI preparing in the background. It's not even really learning in any meaningful sense.
It's a tool. A really powerful statistical tool. And like any tool, it's useful for specific jobs and useless for others.
If you're my age — 24 to 35 — and you've heard about ML but thought it was beyond you, it's not. It's math. It's pattern-finding. It's trial and error dressed up with calculus. You can understand it. Not in five minutes, but in realistic time.
And honestly? In an economy where data is increasingly central to every decision — from which mutual fund to buy (Groww, Morningstar), to whether you get a loan (CRED), to whether your payment gets flagged as fraud (PhonePe) — understanding how these systems actually work gives you a kind of literacy that matters. Not to become a data scientist necessarily. But to not be mysteriously dependent on algorithms you don't understand.
That was worth six months of stupid questions.
If you have questions about any of this, or if something didn't make sense, genuinely — ask. Or tell me what you thought I got wrong. I'm still learning too.
Dattatray Dagale
Data Analyst • Blogger • Mumbai
I'm a data analyst from Kalyan, Maharashtra, working at Morningstar. I write about personal finance, career growth, and everyday life for Indian millennials — the stuff I wish someone had told me earlier.
Written by Dattatray Dagale • 09 June 2026
.png)
.gif)
0 Comments