AI Glossary / Time series and signal processing / Beginner

Delphi method

Definition

The Delphi method is a structured process for collecting and refining judgments from a panel through repeated rounds of questions and controlled feedback. In its classical form, participants submit responses independently and their identities are concealed in the feedback shared with other panel members. A facilitator summarizes responses and reasoning so participants can reconsider their judgments.

What it means

Delphi offers a way to use informed judgment when a decision involves uncertainty, incomplete evidence or developments with little relevant historical data. It is used for forecasting and also for other tasks involving expert or stakeholder views. Those purposes should be distinguished: agreement about a preferred action is not the same as a prediction about what will happen.

The method is more than asking several people for opinions once. Its defining process includes an opportunity to review structured feedback and revise responses. A final summary can represent the panel's collective judgment while retaining disagreement. Agreement does not establish that the judgment is correct or that the panel represents a wider population.

How it works

Specify a clear question, recruit participants with relevant and suitably varied knowledge, and decide how responses will be collected and summarized. Explain the task, information supplied, confidentiality arrangements and planned stopping criteria before the exercise begins.

Collect initial estimates together with reasons. The facilitator prepares balanced feedback, such as a distribution of estimates and anonymized arguments. Participants then submit another response after considering that information. They may retain their earlier position when they judge it justified.

Summarize the final round using the agreed method and report remaining differences, participation and changes between rounds. A mean or median may summarize numerical forecasts, but neither is mandatory for every Delphi design. Modified approaches differ in interaction and feedback arrangements, so describe the actual procedure rather than relying on the method's name alone.

Three practical examples

1. Technology planning: A panel estimates when a clearly defined capability might become operationally feasible. The organization records the assumptions behind earlier and later estimates instead of reporting a single date without context.

2. New service forecasting: Specialists independently assess likely demand where comparable historical data are limited. Structured feedback lets them consider information held by other participants without a dominant meeting voice setting the answer.

3. Research priorities: A group reviews proposed topics over successive rounds. The final result is presented as the participating panel's priorities, not as a statistically representative survey of everyone affected.

Think of it this way

Think of a carefully moderated exchange of anonymous written advice. Each participant first thinks independently, then sees an organized account of others' reasoning and can reconsider. The facilitator's job resembles an impartial editor: clarify and summarize the contributions without turning them into an argument for a predetermined conclusion.

Worked example

In an original numerical illustration, five panelists estimate the number of months required for a precisely defined implementation milestone. Their first-round estimates, in ascending order, are 6, 8, 10, 12 and 24. The median is 10 months, and the range extends from 6 to 24 months, a width of 18.

After receiving anonymized reasoning and a shared clarification of the milestone, the same five submit 8, 9, 10, 11 and 14 months. The median remains 10, while the range width falls to 6. Thus, the central summary is unchanged even though the responses are less dispersed.

This arithmetic does not show that the final estimate is more accurate. The actual completion time is not provided, and the panel's range is not a calibrated prediction interval. Nor should the facilitator discard the 14-month view merely because it differs from the median. Its supporting reason may identify a condition that matters to planning.

Illustrative case study

In a fictional case, Cedar Operations seeks judgments about the rollout time for a new service. Its facilitator discovers that some participants interpret completion as a pilot launch and others as deployment to every branch. She clarifies the milestone before circulating the next questionnaire.

The final report includes the response distribution, the common definition and a remaining disagreement about supplier readiness. Managers use that disagreement to prepare contingency plans. This fictional example illustrates improved clarity and documented reasoning, not a measured improvement in forecasting accuracy. It also shows why apparent disagreement can arise from different questions rather than different expertise.

Common misconceptions and mistakes

Delphi is not an ordinary brainstorming meeting, a one-round poll or a guarantee of consensus. Repeated rounds can still be affected by anchoring, selective feedback, panel selection and participant dropout. Anonymity reduces some interpersonal pressures but does not remove all cognitive or institutional bias.

A narrow spread of responses is not proof of truth. Participants may share the same mistaken assumption or become influenced by the reported group result. Conversely, persistent disagreement may be informative. Stopping only when everyone agrees can suppress useful minority reasoning and place unnecessary pressure on participants to conform.

Business use

The method can make a judgmental process more transparent when quantitative evidence alone does not settle the question. It is most useful when the organization can invest in careful question design, appropriate participants and an impartial facilitator.

Keep forecasts separate from targets and preferred outcomes. Document who participated, how feedback was prepared and why the process stopped. Preserve reasons for disagreement and explain the limits of the panel's coverage. Where outcomes eventually become observable, compare them with the forecasts and learn from the differences. The resulting report should support a decision without implying statistical certainty that the elicitation process cannot supply.

People also ask

Must every participant reach the same answer?

No. The final report can retain meaningful disagreement.

Is a median forecast compulsory?

No. The summary rule should match the task and be specified clearly.

Does anonymity mean the facilitator cannot know participants?

Not necessarily. The confidentiality arrangement should state who can access identities and individual responses.

Can Delphi be used beyond forecasting?

Yes, but consensus about priorities or recommendations has a different interpretation from forecast accuracy.

Check your understanding

What changed in the worked example?

Answer: The range narrowed while the median stayed at 10 months.

Does that prove better accuracy?

Answer: No. The actual outcome is unknown.

Why collect reasons as well as numbers?

Answer: They expose assumptions and help participants evaluate information behind differing judgments.

Related terms

• expert elicitation
• judgmental forecasting
• consensus
• anonymity
• anchoring
• scenario planning
• forecast aggregation
• stakeholder engagement

Learn next

Learn next how to formulate an unambiguous forecasting question, distinguish consensus from calibration, and document uncertainty in expert judgment. Compare Delphi with structured analogies and scenario analysis before choosing a process for a specific decision.

More in Time series and signal processing

• Collaborative filtering - Collaborative filtering is a recommendation approach that learns from patterns of interaction between users and items.
• Cumulative forecasting - Cumulative forecasting predicts the total amount of a variable over several future periods.
• Curve fitting - Curve fitting is the process of choosing parameters of a mathematical function so that the function approximates observed data according to a stated criterion.
• Diagnostic checking - Diagnostic checking is the examination of a fitted model for signs that its assumptions or structure are inadequate for the intended use.
• Difference-equation filtering - Difference-equation filtering computes a discrete-time output from weighted current and past inputs and, in recursive filters, past outputs.
• Differencing - Differencing is a time-series transformation that replaces observations with changes between observations at a specified lag.

Sources

Hyndman and Athanasopoulos, Forecasting: Principles and Practice, The Delphi method: https://otexts.com/fpp3/delphimethod.html . RAND, Methodological Guidance for Conducting and Critically Appraising Delphi Panels: https://www.rand.org/content/dam/rand/pubs/tools/TLA3000/TLA3082-1/RAND_TLA3082-1.pdf . Primary methodological sources were consulted. The panel estimates, arithmetic and organizational case are original illustrations, not reported study findings.

Previous: Deliberative layer Next: Delta encoding