Author: Ron Dembo
Forecasting is a billion-dollar industry ramped into overdrive in response to the 2020 pandemic. Based primarily on historical data, modelling is used now even where it is unhelpful, and it has become a crutch for companies trying to make the best use of their exponentially growing pool of data and their rapidly expanding computing power.
But as the world becomes more complex, with globe-spanning supply chains and worldwide financial webs, and as big data, machine learning, and AI leave us increasingly automated, our vulnerability to radical uncertainty grows and forecasting is proven increasingly ineffective. There is simply too much that must be modelled across systems that are hypersensitive to disruption and rapidly reactive.
Dangerously, forecasting based on historical data tends to increase exposure to risk. It dupes companies into believing that the potential downsides of a decision are within an acceptable range. So where are we going wrong?
You Can’t Predict a Black Swan
One of the biggest mistakes that forecasters make is to believe that they can predict Black Swans. These are events that are outliers, located beyond our ability to conceive because, so far, no evidence can even point to their probable existence. They are also highly impactful despite being so unlikely, so failing to predict them is especially damaging.
Time and again we have proven our inability. There is simply no ideal model that can generalize the specifics of our complex and random reality. For every Black Swan we think we can imagine and account for, another lies in wait unseen. And yet in trying to predict one after the other we waste valuable resources, we neglect other possibilities, and we become more vulnerable.
Past Data Is No Predictor of Future Behaviour
We are enamoured by the precedence set by data, beholden to the belief that it can pinpoint the future.
When forecasters express their visions of the future, they typically do so by assuming that most of the world operates like the past. They assess the risk of a future endeavour by looking back at past data and calculating statistical measures like the standard deviation. Based on the past data, they say, this is a safe bet – no big bumps ahead.
But in real life, variations above seven standard deviations – events which should be impossible in a normal lifetime – happen frequently. They can even exceed twenty or thirty standard deviations, leaving investments in turmoil and forecasters scratching their heads.
This is why forecasting is only useful on a very short timescale, under very specific parameters. It can help us turn the stochastic into the deterministic [Splitting the certain from the uncertain future ] and reduce our uncertainty about the future only in very specific instances.