In September 2019, PMC Bank customers were suddenly told they could take out only ₹1,000 from their own accounts. People lined up outside branches with their passbooks. Everyone asked the same thing: how did nobody see this coming?
That stuck with me. If the warning signs were public, maybe a computer could spot them early. So I decided to try.
The question
Can you tell which Indian bank will get into trouble, using only information that’s free online? By “trouble” I mean an RBI penalty, a SEBI order or a cyber security incident in the next two weeks.
What I did
I collected public information on 25 Indian banks, big and small: RBI notices, SEBI orders, stock exchange filings, court cases, enforcement agency news, cyber security alerts and news stories. That came to about 4,600 documents, each linked to its bank and its original source.
Then I tried seven ways to predict trouble:
- Counting. Are there more documents about a bank than usual?
- Counting by source. Are they coming from the regulator, the courts or the press?
- Counting by type. More penalties? More court cases? More complaints?
- Reading for meaning. A small AI reads the documents, and simple maths looks for warning signs in what it read.
- Word spotting. Do certain words show up before trouble?
- Learning from examples. Show the AI a few past cases and let it learn from them.
- Asking an AI. It reads a bank’s recent documents and answers yes or no: is trouble coming?
Before the final tests, I wrote down the rules and didn’t change them afterwards: what data to learn from, what data to test on, how to score, and when to stop. If nothing did clearly better than guessing, I would stop trying to predict and build something else. I did this so I couldn’t fool myself. When you really want to find a pattern, it’s easy to see one that isn’t there.
The results
None of the seven did better than guessing. Most did slightly worse. That one hurt a little.
Here are the scores, out of 100. Higher is better. Guessing already scores about 26, because trouble is rare. To count, a method needed 31.
| Method | Score (out of 100) | Result |
|---|---|---|
| Just guessing | 25.9 | The score to beat |
| Counting (all three ways) | 24.3 | Worse than guessing |
| Reading for meaning | 26.0 | Same as guessing |
| Word spotting | 25.9 | Same as guessing |
| Learning from examples | 24.5 | Worse than guessing |
| Asking an AI | 24.6 | Worse than guessing |
| Needed to pass | 31.0 | None came close |
Even when I looked at one type of trouble at a time, nothing reached the bar. So I stopped.
I even tried giving the methods the news from the same day the trouble happened. They still couldn’t pick out the right banks.
Would a bigger AI have done better? Probably not. Two Bank of England researchers found that small AI models with simple maths usually beat big AI models when there are only a few examples to learn from, which is the case here.
In short
With public information alone, I couldn’t predict which Indian bank would get penalised in the next two weeks. Not by counting, not by reading, not with AI. Maybe someone will find a way one day. With this data, I couldn’t.
So I built Bank Docket
If I couldn’t predict trouble, I could at least keep a clear record of it. That’s Bank Docket: a free website that shows what regulators have done to each Indian bank, with a link to the source for everything.
It started as 25 banks for the experiment. Today it has:
- 467 banks. Government, private, small finance and payments banks, plus hundreds of district and state cooperative banks. Cooperative banks are where depositors have been hurt most often.
- Nearly 13,000 documents from the RBI, SEBI, CERT-In, the Enforcement Directorate, CBI, PIB, the courts and the stock exchanges. News stories are included too, kept separate from the official record.
- A list of penalties going back to 2014. Each one has the date and a link to the original notice, and the amount when the regulator gives one.
What you can do with it:
- Look up your bank. Is it under RBI restrictions? Is your money insured, and up to how much? What has the RBI done to it in the past year? Each answer links to the official source.
- See official records first. Documents from regulators and courts come first. News stories are in a separate section, so you can tell what a regulator did from what a newspaper said.
- Read what happened at PMC Bank, Yes Bank and Lakshmi Vilas Bank, step by step, from official records.
- Compare up to four banks side by side, or print a one-page summary of any bank.
- Follow a bank. Each bank has its own feed, so you can get updates when something new comes in.
The site updates every night. If a night gets missed, it goes back and picks up any RBI notices it missed. If a source stops updating, the site shows that too.
What it doesn’t do
It doesn’t score, rate or rank banks, and it doesn’t give money advice. That’s because of the experiment above. If the data can’t predict trouble, a score that looks like a prediction would only mislead people.
Bank Docket shows what happened, when it happened and where it’s written down. It shows how up to date its data is, and a limits page lists what it doesn’t cover yet.
Who it’s for
- Depositors who want a quick, clear answer about their own bank.
- Journalists who need a bank’s history with regulators, with sources, quickly.
- Analysts, compliance teams and lenders looking into a bank before working with it.
It’s free and you don’t need an account. Sign up on the site to hear when bank alerts launch.
Prove me wrong
I’d be glad to be wrong about this. If you have an idea I didn’t try, a better way to define “trouble” or data I didn’t have, I’d love to hear it, and I’ll test it the same way.

