Toward a Philosophy of Forecasting

Or, To Forecast is Human

Epistemic status: medium certainty

Why do we forecast? The world is imbued with uncertainty and the human condition is largely one of uncertainty. Forecasting is the means of gaining knowledge about the future, thereby reducing uncertainty. Humans are constantly forecasting, at both micro and macro levels. At the micro level, the level of the unconscious mind, the brain is a prediction machine, constantly making predictions about what is to happen next. At the macro level, the level of human society, there is also an obsession with trying to gain knowledge about the future. There is a high premium being put on getting forecasts about the future, often regardless of whether they’re accurate or not, since they at least seem to reduce uncertainty in the moment. People fight change and hold on to the status quo since change makes the prediction work of the brain more difficult.

Between those two, at the level of the conscious mind, we see the same thing. We suffer from our uncertainty about the future and are therefore constantly seeking knowledge, to reduce the uncertainty. To forecast is therefore very human. A philosophy of forecasting, therefore, should be based on the goal of pursuing knowledge and reducing uncertainty. Some uncertainty is reducible and the way to do so is through forecasting.

When I talk about forecasting, note that I’m in practice referring to superforecasting, as per Tetlock, as opposed to pure finger-in-the-air speculation from political pundits or pure financial extrapolation in finance departments, which is less useful. To have forecasts that truly increase our knowledge about the future, it is necessary to apply the learnings from Superforecasting and leverage the wisdom of the crowd. Other types of forecasts contain very little knowledge. They may only generate information, which is very different. Forecasting in the superforecasting sense are those that contain knowledge rather than just information. They can be said to have a high knowledge quotient, or embodied knowledge potential.

Can we achieve full certainty? Uncertainty is never fully reducible to zero. There will always be black swans and events that cannot be foreseen in any way. Even in the present moment, the fleeting instance where the past meets the future, we can never have full certainty regarding the world.

Leaving the present increases uncertainty. This holds in either direction. Going forward, into the future, the further we go, the less epistemic certainty there is. The same holds for going further back in the past. Memories are fallible, and all history is a perspective, a narrative placed upon discrete events. History in a sense is therefore back-casting, of trying to increase epistemic certainty regarding the past. For forecasting to be possible, however, there must be a past and a future, since there is something there that we are trying to increase our certainty about, which must in some sense already exist.

The level of certainty may be highest in close proximity to the present, but the usefulness of forecasts increases in the other direction. A forecast stating that I will still be typing in my laptop in the next second embodies very little knowledge, while one accurately capturing the state of the world a decade out would have a large knowledge quotient. Early ACE-era Good Judgment forecasts were mostly in the timeframe of less than a year or two out. This year, however, there was an academic tournament called Persuasion, where we forecasted on much longer timeframes. Further research is needed to determine if we can find an “epistemic sweet spot” where the product of forecasting ability and usefulness of the forecast – is optimized. There is also a need to taxonomize different types of knowledge in terms of its differing forecastability, opening up for a whole field of “knowledge studies”.

We should also note that uncertainty seems anecdotally to be on an upward trajectory. This is hard to measure quantitatively, but one attempt – the World Uncertainty Index – shows one measure of uncertainty over the past decades steadily increasing. The Economist also just noted that “the word “uncertainty” appears more than 60 times in the IMF’s latest global outlook, about twice as many as in the April and October 2022 versions.” Since we can assume that all change causes increased uncertainty, there should have been less uncertainty before the industrial revolution, given that the pace of change, by any objective measure, was just slower. Change is largely driven by our technologies, and since technology is combinatory, where the number of permutations constantly increase, so should change, and thereby uncertainty.

Similarly, it would be beneficial if we adapted our language to better incorporate uncertainty. There are many thinkers such as Wittgenstein who have pointed out the inadequacies of language. Even if language is deeply flawed and most information gets lost in communication, some level of knowledge should still be able to be transmitted. It is perhaps true that as C.K. Ogden noted, much of the world’s troubles can be ascribed to the illusion that a thing exists just because we have a word for it. But this should not remove our credence in the epistemic value of all words. And it might be essential to be able to invent new words in order to speculate about their existence prior to potentially bringing them into being. But it would be preferable to integrate levels of uncertainty, and correspondingly, levels of forecasting, into our communication. Scott Alexander does this very well in Astral Codex Ten, where he leads with the level of epistemic uncertainty of his essays. Superforecasting does this well, with its focus on always arriving at a probabilistic forecast with a precise number. A lot of knowledge is compressed into a single number – its level of precision, its distance from the end points of 0 and 100, its distance from the midpoint of complete uncertainty, etc.