Delphi for Forecasting: Expert Prediction Methods

Delphi forecasting is the original use case the method was built for — RAND Corporation developed it in the 1950s specifically to forecast technology and military outcomes, before academic researchers later adapted it for general expert consensus. If your study is asking experts to predict when something will happen, not just agree on what's currently true, you're running a forecasting Delphi, and a few design choices change because of it.

What Makes a Delphi Forecasting Study Different From a Consensus Delphi

Most academic Delphi studies today ask panelists to agree on a current state — best practices, guideline content, definitions. A forecasting Delphi asks something fundamentally different: when will an event occur, how likely is an outcome, or what will a future state look like. That shifts what you're asking panelists to do.

Delphi Forecasting vs. Other Forecasting Approaches

When a Forecasting Delphi Makes Sense for Your Research

A Worked Example

Imagine a panel of 15 experts forecasting when a specific AI capability will become commercially viable. In Round 1, estimates might range from 2027 to 2034, each with a written rationale. In Round 2, panelists see the full spread plus the strongest reasoning from both early and late estimators, and many revise toward a narrower 2028-2031 band — not because they were told to converge, but because the visible reasoning changed their own assessment. By Round 3, if the range stabilizes around 2029-2030, that's a reportable forecast with a documented uncertainty band, not a single overconfident number.

Running a Forecasting Delphi: What Changes in Practice

Instead of "rate your agreement with this statement," ask for a specific value: a year, a probability, a range. Vague forecasting questions produce estimates that are impossible to compare or average meaningfully across rounds.

This is the single most important difference from a standard consensus Delphi. The justification, not just the number, is what gets fed back anonymously and is what actually moves other panelists' estimates in the next round.

Show panelists the full spread of estimates and a sample of the reasoning behind outlier positions — an outlier with strong reasoning should be visible, not averaged away before anyone sees it.

Measure whether the spread of estimates narrows round over round, not whether a fixed agreement threshold is crossed. A forecasting Delphi that never converges is itself a valid and reportable finding — it suggests genuine uncertainty rather than a flawed process.

State the final range or distribution alongside the point estimate, and note explicitly if convergence was partial. A single forecasted number without its uncertainty range overstates what the panel actually agreed on.

Common Pitfalls in Delphi Forecasting

Choosing Software for a Forecasting Delphi

Estimate-based rounds generate messier data than agreement ratings — dates, ranges, and written justifications all need to be aggregated and displayed back to the panel clearly between rounds. A real-time Delphi platform that can show a live distribution of numeric estimates alongside the reasoning behind them removes most of the manual compilation work that makes forecasting studies slow to run with spreadsheets and email.

Delphi forecasting trades the speed of a prediction market for depth: structured reasoning from people with direct expertise, surfaced anonymously, refined over rounds. For long-horizon or data-scarce questions, that trade is usually the right one.