When the future is outside the data
For a long time, most of my questions began with data that had already been observed. I would estimate a relationship, worry about identification, and then ask what the result implied for policy. Econometrics deals in counterfactuals all the time, so the method was never confined to the past. What changed during my PhD was the distance. Some of the futures I was asked to think about sat far outside anything the data had seen.
Distance matters because an estimate is most trustworthy close to the variation that produced it. Push far enough beyond that range and the regression still returns a number, but more of the answer now comes from the functional form and less from anything observed. That is not a flaw in econometrics. It is a boundary, and large policy questions make it easy to see.
The Irish Land Use Review shows the kind of question I mean. Its scenarios include substantial additional forestry and large-scale rewetting of grassland on organic soils by 2050. Those are not year-to-year movements in a land system. They change the structure of the system itself, and no record of past farm decisions was collected under that structure.
Photograph: Pocyun / Wikimedia Commons, CC BY-SA 4.0.
Scenario modelling approaches the problem from the other end. A scenario starts from a target, a policy assumption or a possible condition and asks what else would have to change if that condition held. Backcasting goes further, beginning at the future objective and working back towards the present. The European Commission’s short guide to scenario building in foresight and the OECD’s strategic foresight work draw the same line between exploring plausible futures and predicting one.
Observed behaviour stays essential inside a scenario. If the scenario assumes that farms respond to a price, a payment or a constraint in a particular way, past behaviour is still the best discipline on that assumption. The mistake is letting a historical estimate carry the whole exercise once the system has moved well beyond the conditions that produced it.
Irish afforestation makes the point concrete. Between 2007 and 2025 annual planting ranged from 8,314 hectares in 2010 to 1,573 in 2024. A pathway that holds planting at the current target of 8,000 hectares a year through to 2050 asks for something the recent record has delivered once, in 2010, when farmers did almost all of the planting. A model of planting estimated on the last decade has learned mostly from a decline: farmer planting fell in almost every year from 2014, and the non-farmers who partly replaced them faced different incentives. That model can still say useful things about how planting responds at the margin to a premium or a licensing delay. It cannot say whether farmers would return at several times their recent scale, because nothing in its sample resembles that world. The scenario has to assume it, and an honest report says that it did.
Every scenario result should come with a plain statement of which numbers were estimated and which were assumed. It is a small reporting habit, and it would make many land-use results far easier to read. A reader could then decide how far to trust the extrapolation instead of discovering it in an appendix.
Climate policy has an older version of the same division. An OECD paper on emissions baselines separates top-down economic models, which give economy-wide coherence, from bottom-up technology and sector models, which give detail. I now think about my own methods in similar terms. Econometrics tells me what observed variation can teach. Scenario models ask what a different future would require. Spatial models ask where the change lands. Microsimulation stops me forgetting that thousands of farms facing one policy will not respond in one way.
No single method does all of that well, and I have stopped looking for one that claims to.
Elvis Kwame Ofori
Researcher and writer behind EKO Perspectives.
More from EKO Perspectives · Follow via RSS · About the author