Method, not verdicts
Media, institutions and analysts receive the model and its assumptions, not a number to publish. A forecast whose reasoning cannot be inspected is worth very little.
Elections are one of the few social processes with a hard, public and repeated ground truth: the count. That makes them an unusually honest testing ground for prediction under uncertainty — and a place where a model can be proven wrong in a single night.
Three questions organise the work. First, the gap between stated and actual behaviour: why declared voting intention and the count diverge, and whether that divergence is itself predictable rather than merely noise.
Second, the life cycle of emerging parties: the pattern by which a new political actor breaks through in a low-cost election and then fails to consolidate. It has an observable shape rather than being a sequence of unrelated national accidents.
Third, the voting mechanism itself: what a given allocation rule rewards, what it renders invisible, and how the same ballots produce different chambers under different formulas.
This laboratory does not exist for the electoral business. It exists because elections are the best available testing ground for social prediction. Almost no domain offers a public, repeated, dated ground truth; inside a company, when a forecast misses, one can always argue the target was badly set. Here one cannot.
Working where error is demonstrable forces a more honest method: state the interval beforehand, record the deviation afterwards, and do not rewrite the hypothesis in hindsight. That discipline is what transfers to demand or load forecasting, which is the same problem with a worse ground truth.
Media, institutions and analysts receive the model and its assumptions, not a number to publish. A forecast whose reasoning cannot be inspected is worth very little.
Sessions with the teams who interpret results in real time, focused on telling a real movement from sampling noise before it drives a decision.
The hard part here — inferring behaviour from aggregates and stating uncertainty honestly — is the same problem as demand forecasting. It travels well into industry.
A new party concentrates attention, peaks in a second-order election with a low participation cost, and then fails to hold that support when the electoral arena changes. We are reconstructing that curve across consecutive Spanish cycles to test whether the decay has a stable shape.
The working hypothesis is that the peak measures available discontent rather than support, and that the two behave differently when the cost of voting rises. The interesting part is not the decline: it is that the decline began while media attention was still at its maximum.
Polls are not simply imprecise: they are imprecise in patterned ways, and the pattern differs by party and by type of election. The project builds a record of those deviations to establish whether the correction can be estimated in advance instead of explained afterwards.
A theoretical line on allocation rules: what each system rewards, what it hides and how identical ballots yield different chambers under different formulas. Less applied than the others, deliberately — it is the piece that stops the laboratory mistaking the current rules for the natural order.