Predicting attrition before it happens: the leading indicators

By the time a resignation letter lands in your inbox, the decision is usually weeks or months old. The person has already updated their CV, taken a few interview calls, and quietly made peace with leaving. Predicting employee attrition is rarely about catching the moment someone applies elsewhere. It is about reading the conditions that pushed them there long before, while there is still time to change them.

Most attrition reporting is built backwards. It counts who left, in which team, after how many months. That is useful for an audit and weak for prevention, because every number it produces describes a decision you can no longer influence. The more useful question for an HR or People leader in India, where replacing a mid-level employee can cost a meaningful share of annual salary in hiring, notice-period overlap, and lost ramp time, is different: which teams are drifting toward the exit right now, and what is the earliest honest signal that they are?

A lagging indicator confirms something that has finished happening. Resignation count, regretted-attrition rate, exit-interview themes, and a spike in PF transfer-out requests are all lagging. They are real and worth tracking, but they move after the fact. By the quarter your dashboard shows attrition climbing in a function, the cluster of departures it reflects was usually set in motion a quarter or two earlier.

There is also a measurement trap. Lagging indicators are loudest exactly when it is too late to act cheaply. A team that loses three of its eight engineers in a month forces expensive, reactive decisions: emergency backfills, retention counter-offers, redistributed workload that pushes the survivors closer to the edge. The cost of attrition is not only the replacement spend, which is widely estimated at roughly one-half to two times annual salary depending on the role. It is the second wave of departures that an understaffed, overloaded team tends to trigger.

Leading indicators work differently. They measure the conditions that produce a decision to leave, and those conditions are often visible while they are still reversible. The hard part is that the most honest leading signals are not in your HRMS. They are in how people feel about the work, week to week.

Three families of signal tend to move before resignations cluster. The first is mood: aggregated, self-reported wellbeing check-ins that capture the slow slide from energised to flat to depleted. The second is engagement, read not from an annual survey but from behaviour, such as participation in optional sessions, response to internal communication, and whether people are using the support that exists. The third is sleep and recovery, which is where chronic overload often shows up physically before anyone says the word burnout out loud.

None of these is a crystal ball on its own. A dip in mood during appraisal season or a festival-heavy month can be noise. What matters is the shape over time and the convergence across signals. When a team's average mood drifts down for several consecutive weeks, optional engagement quietly falls, and self-reported sleep worsens in the same window, you are looking at the early form of what the WHO, in ICD-11, describes as burnout: an occupational phenomenon with three dimensions, namely exhaustion, cynicism or mental distance from the job, and a reduced sense of efficacy. Those three dimensions are also, not coincidentally, a fairly accurate sketch of someone halfway out the door.

The mechanism is plausible, not magical. People rarely leave a job they find meaningful and sustainable. They tend to leave when the work has become draining, the recognition thin, and the recovery non-existent, and those states often leave a measurable trace in how a team reports its own wellbeing weeks before anyone formalises a decision.

The instinctive worry is the right one. If you can see that a team is sliding, can you see who inside it is struggling? The honest answer has to be no, and that is a design choice you make deliberately, not a feature you bolt on later. Useful early-warning analytics should operate at the level of the team, the function, or the location, never the named individual. You are reading weather, not following people.

Practically, that means aggregation thresholds: no metric is shown for a group smaller than a set minimum, often around eight to ten people, so no single person can be inferred by elimination. It means individual responses are never exposed to managers or HR, only the anonymised group trend. It means the data exists to prompt a conversation about workload and support, not to build a watchlist. India's Digital Personal Data Protection Act of 2023 sharpens this into law: you need a clear, specific purpose, genuine and informed consent, data minimisation, and you must honour data-principal rights. A wellbeing programme that quietly profiles individuals is not only a trust failure, it is a compliance risk.

There is a performance reason too, beyond ethics and law. The moment employees suspect that a check-in feeds an individual scorecard, the data stops being honest. People manage the metric instead of reporting their state, and your leading indicator quietly turns to noise. Anonymity is not a constraint on the system working. It is the precondition for the data being true.

A leading indicator is only worth collecting if it changes what you do. The point of seeing a team's wellbeing bend in March is to act in April, while the people who matter are still in the building. That action is usually unglamorous: rebalancing a workload that quietly doubled after an unfilled vacancy, fixing a manager relationship, restoring recovery time after a long delivery crunch, or making support genuinely easy to reach rather than three forms deep. The WHO-led estimate that scaled treatment of depression and anxiety can return roughly US$4 for every US$1 invested is encouraging, but the cheaper win is often upstream, in not letting a team reach crisis in the first place.

This is also where the case for investment becomes concrete rather than sentimental. Deloitte's work on the employer cost of poor mental health has reported that prevention and early support tend to carry a positive return, mostly through avoided attrition and recovered productivity. You do not need to oversell any single number to make the decision. You need a leading signal honest enough to tell you where to look, and the organisational will to act on it before the lagging number forces your hand.

Platforms like Aha exist to make that team-level signal collectable and readable without crossing into surveillance, but the discipline matters more than the tool. Track the conditions, not just the count. Read weather, not people. And treat a bending wellbeing line the way a good operator treats a falling barometer: not as a verdict, but as time you have been given to change the forecast.