Alternative Data in Credit: What It Adds and What It Risks

Alternative data in credit is information generated for some purpose other than borrowing, read by a lender to assess a loan. A formal credit record only exists for somebody who has already borrowed, so an absent record is a fact about past access rather than about the person. Alternative data adds a route where none existed, and it risks a signal nobody can see, question or correct.

Most writing on this subject starts with a list of sources and then argues about whether each one works. A list of sources is the wrong end of the problem, and it dates within a year or two besides. Three questions do not date: where a record came from, who can look at it, and what happens when it stops meaning what it used to. The whole subject follows from a single gap, between the group of people a lender would like to serve and the group of people its oldest information source describes.

Why does a lender reach for more information at all?

The pressure comes from the funding side rather than the lending side, and starting there explains a great deal that otherwise looks like curiosity. Rukmini Finance Limited has raised Rs 14,400 crore of borrowings and pays 8.50 per cent a year on them, or Rs 1,224 crore for the stated year. The interest bill arrives whether the money went out of the door or sat where it was. Money raised and not lent earns the institution nothing at all, and costs it 8.50 per cent a year regardless.

So a lender funded in the market needs volume, and volume means assessing people its oldest information source says nothing about. Funding pressure is the honest version of why the subject exists. Nobody in a lending business woke up curious about new kinds of information. The funding side runs on a clock, and the lending side has to keep up with it.

Now the sentence readers do not expect, and the subject keeps returning to it. When a lender cannot assess somebody, the lender simply lends the money to somebody else, so the cost of the failure falls on the person and not on the lender. The Rs 14,400 crore goes out regardless. An application that cannot be underwrittenAssessed and decided on, before the money moves. What the decision settles, and in what order, is covered separately. is not a loss of Rs 14,400 crore to the institution; it is the loss of one loan's margin, and the queue behind it is long. On the other side it is the whole loan.

How big is the stake, in the lender's own money?

Work it on the lender's own stack rather than on anybody's claim. Credit cost is Rs 396 crore for the year, or 2.20 per cent of assets under management of Rs 18,000 crore. Profit before tax for that same year is Rs 450 crore, or 2.50 per cent when it is set on that same Rs 18,000 crore. The two lines are startlingly close in size: the credit cost line is 88.0 per cent as large as the profit line.

So take a movement of a tenth of a point. Move credit cost from 2.20 per cent of assets under management to 2.30 per cent of the same base, and the charge goes from Rs 396 crore to Rs 414 crore. The increase is Rs 18 crore. Set Rs 18 crore against profit before tax of Rs 450 crore and it is 4.00 per cent of it. Set the same Rs 18 crore against assets under management of Rs 18,000 crore and it is 0.10 per cent of it. Set it against the year's interest bill of Rs 1,224 crore and it is 1.47 per cent of that. One amount, three bases, three completely different impressions. A figure quoted without its base has said almost nothing.

One sentence has to come early, before anything after it can be misread. A tenth of a point moves four per cent of the profit, and that is exactly why a lender reaches for more information. Whether a particular source of information moves that line, in which direction and for whom, is a question only a lender's own approved loans and its own declines could answer. The size of the stake can be worked from published lines; the effect cannot.

What does a formal credit record actually record?

The block that reframes everything after it comes now, so take it slowly. A credit record is a record of loans that were made, and of how those loans were repaid. Read that sentence again with the emphasis on the first half. Loans that were made.

A credit record is therefore a record of access at least as much as it is a record of conduct. Somebody who was never lent to has no record of repaying, and that absence is a fact about what was offered to them rather than a finding about them.

A shopkeeper's book makes the shape plain. He writes down who owes what, and he adds a name to the book on the day he first gives that person goods on credit. Every name in the book got there because he decided, once, to extend credit. A blank leaf in that book says something about the book. The blank leaf says nothing about anybody outside the book. The framing is mechanical rather than sympathetic: the process that generates a credit record begins with a lending decision, so anybody who was not lent to is missing by construction.

Who holds the record, and how lenders report into it, belongs to a credit information companyThe entity that collects loan and repayment entries from lenders and produces a credit score from them. What it may hold, how it is reported to and how the score is built are all covered separately. and is covered separately. Only the construction matters at this point. Both lenders in this record feed it the same way: Suvarna Commercial Bank Limited on advances of Rs 1,44,000 crore, Rukmini Finance Limited on a book 8.0 times smaller at Rs 18,000 crore, and neither of them contributes a single entry about anybody they declined.

What counts as alternative data, and what does not?

Define it by origin, not by a list. Alternative data is information that came into existence for some purpose other than assessing a loan, and is then read for that purpose. The test is one question, and it is cleaner than any list of sources could be: ask what the record was created for.

Here is the test working on something everyday. Setu Payments Limited carried 1,200 crore transactions worth Rs 3,60,000 crore in the stated year, an average payment of Rs 300.00/-. Every one of those records exists because Rs 300.00/- or thereabouts had to move from one place to another. Not one of them was created to help anybody decide a loan. Apply the origin test and the answer is immediate, and it will still be the right answer in ten years. A list of sources written today will be quaintly out of date long before then.

Four kinds, named by what they are rather than by any product, platform or provider. Records of money moving through an account the borrower has agreed to share. Records of payments made regularly for services. Records of a small business's own sales. And records generated by the channel through which the borrower reached the lender, usually a lending service providerSomebody who finds the borrower, runs the paperwork or chases the money on a lender's behalf, while the loan itself stays with the lender. What is required of one is routed below rather than stated. rather than the lender itself. The fourth is the one that carries the most argument, and it belongs on the list for that reason rather than in spite of it.

What does it add, and who does it add it for?

Three things, and the block states for whom in the same breath as each.

Alternative data adds a route where none existed. Somebody with no formal record can be assessed on something rather than declined for the absence of everything. The first claim ends there, and it is a large one.

Alternative data adds recency. A record of money moving this month describes this month. A record of loans made and repaid describes whatever happened when those loans were live. For a lender pricing something today, the difference in vintage is real.

Alternative data adds granularity. The average payment of Rs 300.00/- is the case in point. A pattern spread across a great many small amounts carries more information than one summary number does, in the same way that a month of a vegetable seller's daily takings says more about the stall than the monthly total does. The total hides every good week and every terrible one.

And now the honest limit, standing in the same block as the claims rather than in a footnote after them: none of this is a claim that assessment becomes accurate, only that assessment becomes possible where it previously was not. Possible and accurate are different sentences, and the gap between them is where most of the overselling on this subject lives.

What can a signal stand in for without anybody intending it to?

The first of the three costs is also the least visible of them. A variable that predicts repayment in a sample may be standing in for something else entirely, and the lender need not know what.

Where somebody lives, what work they do and what they can afford are all correlatedMoving together, so that knowing one says something about the other. It says nothing about which causes which, or whether either causes the other at all. with almost everything, so a signal can carry information about all three without anybody choosing it for that. Nobody sat down and decided to key a lending rule off a postal code. But a postal code travels with the price of housing. The price of housing travels with income, income travels with occupation, and occupation travels with a dozen other things nobody listed. Choosing one quietly chooses all of them.

So a rule built on such a signal can produce a pattern of outcomes nobody designed and nobody inspected. The mechanism holds however often it happens, and how often it happens is measured nowhere in this record. The fair practice requirements a lender works under, and what it must be able to explain to a borrower about a decision it has taken, are set by the Reserve Bank of India at rbi.org.in and they move.

What happens when information was generated for a different purpose?

Information generated for one purpose and read for another is the entire definition of this subject. The reuse is also the central difficulty, and the two facts are the same fact seen from two sides.

Agreeing to share a record is not the same as understanding what will be read out of it, and the second is what makes agreement mean anything. Somebody can know exactly which record is being handed over and have no idea at all what will be inferred from it. The inference was not in the record and did not exist until somebody built a rule. The gap between handing over a record and knowing what will be read out of it is not closed by a longer form. The gap closes, if it closes at all, when somebody is told what the rule looks at and why.

A household version. A customer hands a shop a telephone number so the shop can call when the repair is done. The call has been agreed to. Being telephoned every week for a year about something else has not been agreed to, and the fact that the number is the same number in both cases is exactly why the distinction has to be made out loud rather than assumed.

The Reserve Bank of India at rbi.org.in sets what a lender may collect, keep and share about a borrower, and on what consent, and that requirement moves. The routing table further down names where that requirement is settled, along with the five others this subject touches.

Can a person see an alternative signal, and can they correct it?

The third cost, and the sharpest of the three. A formal credit record can be seen by the person it describes, and there is a defined route for putting an entry right. How an entry is corrected, and by whom, is set by the Reserve Bank of India at rbi.org.in.

An alternative signal frequently has no such route. There may be nothing to look at, nothing to point at, and nobody to point at it to. A person may never learn that a decision turned on a particular signal, and having never learned it, has nowhere to take an error in it.

The two kinds of information differ in a way that has nothing to do with accuracy: whether the person described can see the information and put it right. The sentence worth carrying away has little to do with which kind predicts better. Standing, not quality, is what separates them.

Does a relationship found in one period hold in the next?

A relationship measured in one period need not hold in the next. Nothing guarantees that it will, and nothing announces when it has stopped. A lender learns about a relationship that has stopped working only from the loans it approved. The narrow window is what makes the problem worse than it sounds.

A signal that quietly stops working produces losses on the approved bookThe loans an institution actually made, which is the only population it ever observes repayment for. The consequence of that is covered separately. and declines on everybody else, and it sees only the first of the two. Half the consequence arrives as a number in the accounts. The other half arrives as nothing at all, permanently, and no amount of care applied to the first half will ever surface the second.

One year is on record, with no second year and no series, so only the mechanism can be stated. A drift figure would need a series, and a made-up series showing a signal decaying would look like evidence while being nothing of the kind.

How would a practitioner read all this on a real institution?

Somebody outside a lender cannot measure whether its information sources work. The question needs the lender's own approved book and its own declines, and even the lender has only half of that. So the questions worth asking from outside are of a different shape, and there are four of them.

First, the stake. The arithmetic is one anybody can do from published lines. At Rukmini Finance Limited a tenth of a point of credit cost on assets under management of Rs 18,000 crore is Rs 18 crore and 4.00 per cent of profit before tax of Rs 450 crore. The sum, done on whichever institution is under examination before a word of its commentary is read, says how much any of this is worth to that institution and therefore how hard it is likely to be pushing.

Second, the treatment a rule gives an empty field. One sentence of an answer settles it, and the answer is diagnostic, as the failure block below sets out.

Third, whether the institution can say what a decision turned on. Not the model, covered separately and in full; simply whether an explanation exists at all and who is entitled to it.

Fourth, where a person takes an error. If the answer is a formal credit record, the correction route is set by the Reserve Bank of India at rbi.org.in. If the answer is a signal with no route, that is a real difference and it should be named as one. A practitioner reading this subject well spends almost no time on which sources a lender uses and almost all of it on how the lender behaves when a source says nothing.

Two things worth fixing before anybody starts reading meaning into small movements. The first is a ceiling, and it is a ceiling rather than a forecast. Net interest income of Rs 1,386 crore less operating expenses of Rs 540 crore leaves Rs 846 crore, so a credit cost of Rs 846 crore on assets under management of Rs 18,000 crore, or 4.70 per cent of that base, would take profit before tax to exactly nil. The ceiling is not a destination the line is heading for; it marks how far the line could travel before there is nothing under it. The second is about precision. Suppose the movement in credit cost wanted is the one costing exactly 5.00 per cent of profit before tax of Rs 450 crore. In whole crore there is no such amount anywhere between nil and that ceiling: Rs 22 crore prints 4.89 per cent, Rs 23 crore prints 5.11 per cent, and nothing between the two is a whole crore. A share quoted to two decimal places is finer than the amounts that produced it.

Why a control over approvals would teach nothing

The control such a subject would ordinarily carry is the amount or kind of information used, set against the number of applications approved. Two grounds count against building one.

The first ground is editorial: the picture such a control draws is a picture of somebody being refused credit, moving under a reader's finger, and it teaches nothing the arithmetic in the text above does not already teach. The Rs 18 crore against Rs 450 crore division stands in the text above, in words, where anybody can check it. Wrapping it in a slider would add an experience and subtract nothing from the difficulty, and the experience is one of treating somebody's access as a dial.

The second ground is evidentiary: this record contains no measurement of what any source of information contributes to any outcome, so every position of such a control would be an invented figure dressed as an illustration. A number that a reader can move is a number a reader believes, and a made-up number that moves is worse than a made-up number that sits still.