Delphi Technology Forecasting

Generic Delphi can seek consensus on any question — a policy preference, a definition, a priority ranking. Technological Delphi specializes the method to forecasting: it asks experts to predict the timing or probability of well-defined technological events, such as the year a capability will be achieved or the chance it will exist by a given date. The outputs are therefore numerical forecasts — converged median dates and event probabilities — rather than agreed statements, and the design effort goes into specifying events precisely enough that their occurrence could later be verified. This dating-of-events focus is the original RAND application that gave Delphi its reputation.

Forecast-year distributions are typically skewed, with a few experts predicting much earlier or much later than the rest, and a mean would be pulled toward those extremes. The median gives a robust central estimate that is not distorted by outliers, and the interquartile range — the spread of the middle half of the panel — gives an interpretable measure of disagreement. Reporting these back as feedback focuses each subsequent round on where genuine uncertainty lies. It also lets the facilitator define a clean convergence rule: stop when the interquartile range for the events of interest has narrowed below a set threshold.

No. Convergence means the panel has reached stable agreement, which is valuable but is not the same as accuracy. A well-composed, well-informed panel converging with reasoned justification is more trustworthy than a single guess, yet a homogeneous panel can converge confidently on a wrong answer through shared blind spots. For this reason good practice preserves and reports the residual spread and the dissenting rationales rather than discarding them, treats the median as a best estimate rather than a certainty, and triangulates the Delphi result against quantitative methods such as growth-curve or trend-impact forecasting.