A system can still be operating, a company can still be profitable and a network can still be carrying its load while a deeper problem is already developing. The difficult part is not recognising a crisis after it has arrived. It is deciding which small deviations are meaningful and which are ordinary noise.
In the photograph, the leaf is still mostly green. It has not collapsed, yet numerous damaged patches are already visible across its surface. A single image is not enough to diagnose the plant, but as a picture of a complex system it captures an important stage: the problem is visible long before complete failure.
Human systems often behave the same way. Companies rarely wake up bankrupt without a history. Power grids may accumulate vulnerable nodes before a major outage. Financial systems can look stable while debt, concentration or fragile funding structures quietly grow underneath. The useful question is therefore not “which tiny problem predicts a disaster?” but when do small deviations stop looking random and begin to show that the system is losing its ability to recover normally?
A system can weaken before it fails
Research on complex systems uses the idea of critical slowing down: as some systems approach a critical transition, recovery from small disturbances can become slower. A resilient system absorbs a modest shock and returns to normal; a less resilient one carries the effects for longer.
A landmark Nature review showed that researchers have studied such early-warning behaviour across very different systems, from ecosystems and climate to financial markets. Before some critical transitions, variability can increase, the present state can become more dependent on the previous state, and recovery from small disturbances can slow.
This is not a universal alarm that flashes before every crisis. Newer research continues to develop warning methods for high-dimensional systems, but identifying the exact turning point remains difficult. A signal matters not simply because it exists, but because it repeats, spreads and changes the system’s capacity to recover.
One bad number is weaker evidence than a change in system behaviour
Imagine a company whose sales fall 15% in one month. Seasonality, a delayed contract or random variation could explain it. Now add other facts: sales have fallen for three consecutive months, existing customers buy less often, payment times are lengthening, finished inventory is rising and the company needs deeper discounts to maintain volume. Each observation still has alternative explanations, but together they begin to describe a mechanism.
This is why modern bankruptcy prediction rarely relies on a single magic ratio. A 2024 study comparing failure-prediction models used a large company sample and dozens of financial and macroeconomic variables. The practical lesson is broader: a weak signal becomes more informative when independent indicators start telling the same story.
Financial supervision already works this way
Banking regulation offers a practical example of early-warning logic. Current BIS guidance requires banks to monitor sets of liquidity indicators that can reveal rising vulnerability, including unstable funding, concentration and currency mismatches.
The logic is simple. If supervisors wait until a bank can no longer meet its obligations, an early-warning system is no longer useful — the crisis is already a fact. The useful indicators are the ones that show how a system is becoming more vulnerable before the final outcome occurs.
A small failure becomes dangerous because of where it occurs
Counting problems is not enough. Ten damaged peripheral elements can be less dangerous than one damaged critical node. In complex networks, a local overload can shift stress onto neighbouring components and trigger further failures. A Nature review on network resilience emphasises that systemic risk depends not only on how many elements fail, but also on network structure and adaptive capacity.
That is why “how many problems are there?” is often less useful than “where are they, and what depends on those points?” Losing one ordinary customer may mean little. Losing a customer responsible for half of revenue changes the structure of risk. One resignation is normal; losing the only person who understands a critical system is different.
The most important warning may be slower recovery
Healthy organisations have bad months and recover. Digital services suffer incidents and return. Supply chains encounter delays and reroute. Concern should rise not only when the shocks get bigger, but when the same kind of shock leaves a longer mark than it used to.
Sales take longer to recover after every weak month. Infrastructure incidents take longer to fix. Reserves are being consumed faster than they are rebuilt. Employees spend more time firefighting and less time on normal work. Each event may still look small, yet together they suggest that resilience is being depleted.
How to separate a warning signal from noise
Four properties are particularly useful. First, repetition: a one-off deviation carries less information than the same deviation returning under similar conditions. Second, spread: a problem moving across departments, regions or nodes is harder to explain as a local accident. Third, slower recovery. Fourth, confirmation from independent indicators.
The strongest signal is often not an extreme value in one metric, but several different metrics moving in the same direction. Early warnings should therefore change the level of attention before they change the conclusion: we may not yet know whether a crisis is coming, but we may already know where closer investigation is justified.
Sometimes a spot is just a spot
If every deviation were treated as the beginning of a major problem, early-warning systems would create chaos of their own. Companies would constantly change strategy, investors would react to every fluctuation and engineers would redesign systems after every minor fault. Some large failures also arrive as external shocks with little internal warning.
The goal is not to predict everything. It is more practical: identify the point at which the probability of a structural problem is high enough that checking costs less than ignoring it.
A small sign matters for what it changes, not for its size
One damaged patch on a leaf does not tell us its future. Even many patches do not reveal the cause without context. But they create a question that would not exist on a uniformly healthy surface: is there an ongoing process here?
One weak month is not a crisis. One outage does not prove infrastructure decline. One resignation does not mean a team is collapsing. Yet when deviations repeat, spread, receive confirmation from independent indicators and leave increasingly long recovery periods, the quality of the evidence changes.
Small signs matter not because they reveal the ending. They matter because they give us a chance to ask the right question before the ending becomes obvious to everyone.
