Early Warning Indicators: Triggering Action Before the Limit

An early warning indicator is two things fastened together: a trigger set in advance and an action that starts when it is crossed. Remove the action and it is a measurement. Remove the trigger and it is a discussion. Its whole value is the distance between the trigger and the thing it warns about, and at Vindhya Commercial Bank Limited, invented, one of the seven warns Rs 768 crore ahead of its limit.

Almost every institution that says it has an early warning system actually has a set of numbers that change colour. The numbers are usually right. The colours are usually calculated correctly. Nobody ever wrote down what was supposed to happen, so when the thing they were meant to warn about arrives, nothing does. The missing sentence is what separates a warning from a wall of numbers, and one measurement shows whether that sentence would have arrived in time.

What is an early warning indicator, and what makes it different from any other measure?

The mechanism is identical in an ordinary household, where the rupees are smaller. A household running on one salary looks at the bank balance on the twentieth of the month. Looking at the balance is a measurement. Now suppose the same household has agreed, out loud, in advance: if the balance on the twentieth is below Rs 20,000/-, the eating out stops until the salary lands. The agreed rule is a different object. Nobody has to decide anything on the twentieth. The decision was taken weeks earlier, at a calm moment, by people who were not yet worried, and all that remains on the day is to read a number and do what was already agreed.

An early warning indicatorA trigger set in advance with an action attached to it. Crossing the trigger starts something without a decision being taken in the moment. is that second version, scaled up. An indicator is not a kind of measurement at all: it is a promise, made in advance, that if a stated quantity reaches a stated place then a stated thing will happen. The measurement is only how the promise gets evaluated. The distinction between a promise and a measurement sounds pedantic until an institution turns up that has built a wall of measurements and thinks it has built a warning system. At that point the distinction is the entire difference between the two.

So the object has exactly two parts. The triggerThe stated level at which the indicator fires. It must be evaluable from the document that defines it and not from a discussion. is the stated level. The attached actionWhat happens when the trigger is crossed, written in advance. Writing it in advance is what separates a warning from a measurement. is what starts when the level is crossed. Trigger and action are written down together, in one place, before anybody is under pressure, and neither survives being separated from the other.

What are the two parts doing, and what breaks when either one is missing?

The removal of the action is the common failure, and it is almost invisible. Consider it first. The measurement survives. It still gets computed, it still gets a colour, it still appears in a paper in front of a committee. The sentence saying what happens next has gone, and its absence looks exactly like its presence right up until the day it matters.

At Vindhya Commercial Bank Limited one single measurement appears on both of its lists at once, and no demonstration is cleaner. The share of its certificates of deposit that get rolled at each auction is early warning indicator W2, with an amberThe first trigger level, meaning something starts, and useful only in proportion to how far it sits from the limit. level below 90 per cent, a redThe second trigger level, and a red set at the limit reports a breach rather than warning of one. level below 75 per cent, and an action fastened to it. The same roll rate is also one of the sixteen key risk indicators on the bank's monthly dashboard, where it reports a position to a committee and nothing is written down about what follows. One measurement, computed once, sitting in two places, and the only difference between them is a sentence.

Now take the trigger away instead and keep the intent. The second failure sounds responsible in the minutes. A committee agrees that the funding position is uncomfortable, that it will be kept under close review, and that action will be taken if things deteriorate further. Every word of that is sincere. None of it is an evaluable triggerA trigger that can be judged crossed or not crossed by reading the document, without anybody exercising judgement.. Deterioration has no level. Nobody holding only that minute can ever say, on any given day, whether the trigger was crossed.

The test of a trigger is brutally simple: hand the document to a person who was not in the room and ask them to answer yes or no. If they can, the document holds a trigger. If they have to ask what was meant, the decision has been moved back into the meeting. The meeting is precisely the place the trigger was written to keep it out of. A commitment to act on deterioration is a promise to hold a discussion later, and a discussion held later is held by people who are by then frightened, tired and short of time.

How wide does the band between the warning and the limit have to be?

Here is the question almost nobody asks, and it is the one that decides whether the indicator is worth building. An early warning indicator warns about something. The something warned about is usually a limit, a cap or a level at which the institution is in trouble. The bandThe distance between the warning level and the limit it warns about. It is the entire amount of warning the indicator can give. is the distance between where the indicator fires and where the trouble sits, and that distance is the entire quantity of warning the indicator is capable of producing. Not most of it. All of it.

The household again. The rule fires when the balance drops below Rs 20,000/- and the rent of Rs 18,000/- goes out on the twenty fifth. The band is Rs 2,000/-, and put that way the rule is obviously useless: any ordinary week of groceries clears it. Nobody in that household is careless. The household simply never converted the trigger into the unit that the events actually arrive in.

The conversion into rupees is the whole craft. Transactions do not arrive in percentage points; they arrive in rupees, so a band stated in percentage points says nothing. One percentage point sounds like a comfortable margin. On Vindhya Commercial Bank Limited's deposit book of Rs 76,800 crore, one percentage point of deposits is Rs 768 crore, and that is a number that can be held up against the size of a single deposit to form a view about.

So hold it up. Breach B4 in this bank was crossed in month 10 when a single state undertaking placed Rs 1,440 crore. Rs 1,440 crore against a band of Rs 768 crore is 1.875 times the band. One ordinary placement by one customer is nearly twice the entire quantity of warning the indicator was built to give.

What does a real set of seven early warning indicators look like?

Vindhya Commercial Bank Limited runs seven of these in its liquidity work, numbered W1 to W7, and every one of the fourteen trigger levels below is the bank's own invention. None of them is a supervisory figure. Each is a level the bank settled on for itself. The levels are not what makes the set worth reading. Reading them against the bank's own limit set immediately turns up something the bank does not appear to have noticed.

The bank itself has not taken the next step, and taking it changes how the set reads. Now convert every one of those bands into rupees. Five of the seven convert cleanly on the bank's own quantities. W1 converts into a cost rather than a stock. W3 does not convert at all: its band is two weeks of time, and time is not a rupee amount.

What do the seven feed, and what happens at month 12?

An indicator that fires has to fire into something. In this bank the seven feed the three stageA defined level of response in a plan, entered when stated indicators fire, and covered separately. levels of its funding contingency plan, numbered P1 to P3. P1 is heightened monitoring, triggered by any two early warning indicators turning amber. P2 is a name-specific stress. P3 is a market-wide stress. The plan's contents, what its actions raise and how much of it survives a real stress are covered separately. Only one thing matters for an indicator here: a defined thing exists at the far end of the trigger. An indicator that fires into nothing has no attached action and therefore is not an indicator.

At month 12, the record locks two of the seven as triggered: W4 is red and W7 is amber. Two of seven is 28.6 per cent, and this is a point that calls for care in how it is stated. The fraction 28.6 per cent turns up three separate times in this one bank, on three unrelated things. Two of the seven early warning indicators are triggered. Forty two of the 147 risk data elements carry all eight of their attributes, again 28.6 per cent. And of the seven governing bodies other than the board itself, two carry independent directors, 28.6 per cent once more. A bare 28.6 per cent in this institution means nothing at all until somebody names which set it was drawn on.

And now the awkward part. Stage P1 has to be evaluated from the document that defines it. P1 is triggered by any two early warning indicators turning amber. At month 12, W4 is red and W7 is amber. Is that two ambers or one? If red counts as having passed through amber on the way, it is two and the stage is triggered. If the trigger is read exactly as written, only one indicator is amber and it is not. Nothing in the record settles it.

Do not read that as an accusation of sloppiness. The trigger is a perfectly sensible sentence, written by sensible people, and its weakness only appears in a state of the world nobody imagined while writing it. The general shape of the failure is worth more than the specific instance: a trigger has to be evaluable in every state, not only in the state its author had in mind. The cheapest way to find these is to take the trigger sentence and try to answer it against every combination of statuses the indicators can actually show. The exercise takes an afternoon, and it is the only test that finds this class of problem before the day it matters.

What does indicator W4 actually warn about, and how much room is in it?

W4 is the row that teaches. Read it slowly. Its red trigger is 12.0 per cent of total deposits. Limit L12, set by the board risk management committee, says the top twenty depositors do not exceed 12.0 per cent of total deposits. The red trigger and the limit are the same number. An indicator whose red trigger sits exactly on the limit is not warning of a breach; it is reporting one, in a colour.

Which means the entire quantity of warning W4 is capable of giving lives in the gap between its amber at 11.0 per cent and its red at 12.0 per cent. The gap is 1.0 percentage point of a deposit book of Rs 76,800 crore, or Rs 768 crore. Everything W4 can ever do for this bank happens inside Rs 768 crore of deposits, and once the top twenty pass Rs 9,216 crore the indicator has nothing left to say that the breach log will not say by itself.

There is a second thing hiding in those two levels, and it is the sort of coincidence worth noticing rather than worth building on. The amber sits at 11.0 against a red of 12.0, so the warning fires when the position reaches eleven twelfths of the limit, or 91.7 per cent. The same 91.7 per cent appears as 13,200 over 14,400: the share of this bank's own liquidity buffer that its own severe scenario consumes in thirty days. Two entirely unrelated things in one institution landing on one fraction. The coincidence proves nothing. Recognising it is what keeps one from ever being treated as evidence about the other.

The band is a design choice, and the choice is visible as a single falling shape. Push the amber trigger towards the limit and the band closes; pull it away and the band opens, at the price of firing more often on movements that turn out to be nothing. The six settings below are worked out in full, and the middle one is the point where the band is exactly the size of the single deposit that caused the breach.

How far is the survival horizon from its own triggers, in rupees?

W7 makes the same argument from the other side, and it is the more useful one to have seen. Its band looks generous in its own unit and only becomes legible once it is priced. W7 watches the survival horizon under the bank's own severe scenario, turning amber below 40 days and red below 30. At month 12 the horizon is 34 days. Four days from amber, six days above red. Nobody sitting in a treasury can convert a day into an instruction, so four days and six days are almost impossible to act on.

Price them and they become instructions. Under that scenario the bank's own modelled net outflows run at Rs 720 crore a day on days 1 to 5, Rs 600 crore on days 6 to 10, Rs 420 crore on days 11 to 20, Rs 240 crore on days 21 to 30 and Rs 300 crore a day after that. Cumulative outflow to day 30 is 3,600 plus 3,000 plus 4,200 plus 2,400, being Rs 13,200 crore. The buffer is Rs 14,400 crore. So reaching day 30, the red trigger, needs Rs 13,200 crore and the bank is Rs 1,200 crore above it. Reaching day 40, the amber trigger, needs 13,200 plus ten days at Rs 300 crore, being Rs 16,200 crore, and the bank is Rs 1,800 crore short of it. Rs 1,800 crore of extra buffer is something a treasurer can go and do something about in a way that four days is not.

Can a band ever be too narrow to help, whatever level it is set at?

Yes, and this is the case that widening the band does not fix. The instinctive repair is always the wrong one, so the case is worth meeting properly.

Who actually uses one of these, and what do they do with it?

Three different people pick up an early warning indicator and use it for three different things, and it is worth seeing all three because the object looks different from each seat.

The first is the person inside, the treasurer. Devendra Achar, who carries the risk ownership of breaches B3 and B4 at this bank, does not read W1 to W7 for the position; he already knows the position. He reads them to learn what he is now committed to doing. An attached action written in advance earns its keep on exactly that morning. When the roll rate on his own certificates of deposit comes in below 90 per cent, a version of himself in a calmer month has already settled whether the situation warrants a response, so he is not arguing about it. The advance action is not a constraint on the treasurer; it is a decision made by the treasurer at his most clear headed, delivered to himself at his least.

The second is the analyst outside. Somebody assessing this bank as a counterparty cannot see W1 to W7, but they can do the test the bank did not: taking every disclosed limit and every disclosed warning level and checking whether any warning level sits on top of a limit. The check costs ten minutes, and it reveals something about the institution rather than about the number. An indicator set where the reds have quietly been aligned to the limits is a set that will produce colours on the way to a breach and no time in front of one.

The third is a person with no bank at all. A tea stall outside one office building reorders milk when the stock falls to ten litres, and sells forty litres on a working day. Ten litres is a band of a quarter of a day against a supplier who takes half a day to deliver. Ten litres also feels like a sensible amount of milk, and the stall runs out perhaps twice a month without ever understanding why. Price the band in the unit the events arrive in and the answer is instant. Every early warning indicator ever built, at any scale, is judged by the same single question: is the distance between the trigger and the trouble bigger than the event that will cross it?

What does an early warning indicator not do?

An early warning indicator does not decide anything. It carries out a decision that was already taken. The quality of an indicator is therefore almost entirely the quality of the thinking that went into it months earlier. An indicator also cannot detect what it is not watching: one set on the depositor share sees the depositor share and nothing else, and the whole art is in choosing the quantity that moves first rather than the quantity that matters most. The two are frequently not the same. In this bank the quantity that matters is the concentration and the quantity that moves first is the pipeline of placements being negotiated, so the two are demonstrably not the same.

And it is not a forecast. Nothing in W1 to W7 predicts anything. Each one states a level and a consequence, and the only claim being made is that if the level is reached then the consequence starts. A conditional consequence is a much smaller claim than a prediction, and it is the reason the object works at all: a promise about what the institution will do is entirely within its control in a way that a statement about what the world will do never is.