The Delphi Method in Demand Forecasting

Put five experts in a room and the loudest one wins. Not the most accurate one, the loudest, and everyone in the room knows it while it's happening. The Delphi method fixes that by keeping everyone anonymous until the numbers agree, and it's been quietly doing so since the 1950s.

The Delphi method is a qualitative forecasting technique that gathers anonymous input from a panel of experts over several rounds, sharing summarized results between rounds until the group converges on a consensus forecast. The anonymity strips out groupthink and dominant voices, which makes it useful when there's no data, like a new product or market.

Key takeaways

• Anonymity is the whole mechanism: estimates carry no names, so seniority and charisma stop outranking evidence.
• It's rounds, not a meeting: estimate, see the group's summary, revise, repeat until the spread narrows.
• Convergence is earned, not forced: holdouts explain their reasoning, and sometimes the outlier's argument moves the group instead.
• It's built for data-free decisions: genuinely new products, new markets, and long-horizon calls where history can't help.

What is the Delphi method?

The Delphi method is a structured way to forecast by collecting anonymous expert estimates across rounds and feeding back the results until they converge. Developed at RAND in the 1950s for long-range technology forecasting, it survives because it solves a permanently human problem: groups contain more knowledge than any member, and group dynamics routinely prevent that knowledge from surfacing.

A structured, anonymous panel technique

Three design choices define it. Anonymity: estimates are submitted privately and shared without names, so ideas compete instead of statuses. Iteration: multiple rounds let people update on the group's reasoning rather than its hierarchy. Controlled feedback: a facilitator shares the summary (the range, the median, the arguments), not the noise. Together they turn "a bunch of opinions" into something closer to a measurement of what the panel collectively knows.

How do the consensus rounds work?

Each round, experts submit estimates, a facilitator summarizes the spread, and everyone revises, repeating until the panel lands on a stable number.

The round-by-round process

• Round one: each expert submits an estimate (ideally a low/likely/high range) plus one sentence of reasoning, privately.
• Feedback: the facilitator circulates the anonymous summary: the median, the spread, and the key arguments at each end.
• Round two: everyone re-estimates, having seen the group's reasoning. Outliers either move toward the center or explain why they won't.
• Repeat until stable: typically two to four rounds. Convergence means the spread stops narrowing, not that everyone matches exactly.

A concrete shape: five estimates for a launch open at 800, 1,500, 2,000, 2,400, and 4,000 units. After the round-one summary (and the 4,000's reasoning turning out to assume a retail deal nobody confirmed), round two lands 1,400 to 2,300. Round three settles around 1,800 with a spread the group accepts. The number matters less than what happened to the 4,000: it wasn't shouted down; it was examined, and its hidden assumption surfaced. That's the method working.

Why anonymity matters

Groupthink isn't a character flaw; it's a predictable response to visible hierarchy and social cost. The junior analyst with the pessimistic (and correct) number stays quiet in a conference room and submits it happily in a Delphi round. Anchoring dies too: nobody's first number becomes the room's gravity, because nobody hears a number before writing their own. What survives is the panel's information, which was the point of assembling a panel.

When do planners use the Delphi method?

Planners reach for Delphi when data is missing or unreliable: new launches, new categories, or major market shifts where history can't guide you.

Best-fit situations

• Genuinely new products: nothing comparable in the catalog, so structured judgment is the forecast.
• New markets or channels: experience exists on the panel (sales, partners, advisors) but not in the sales data.
• High-stakes, low-data bets: a large buy where one dominant voice steering the room would be expensive.
• Long horizons: category-level calls a year or more out, where extrapolation is theater anyway.

Its cost is speed: real rounds take days, so it's poorly suited to weekly reorders and perfectly suited to the handful of decisions each year that are big, uncertain, and data-free. For everything lighter, the same principles compress into simple habits (independent estimates first, ranges, written assumptions), which is the broader toolkit of qualitative forecasting.

Frequently asked questions

How many experts do you need for the Delphi method?

Classic practice runs five to twenty. Below five, one person's blind spot dominates; beyond twenty, coordination costs rise faster than added insight. Diversity of vantage matters more than headcount: a buyer, a marketer, a founder, and a supplier see different halves of the same demand.

How many rounds does the Delphi method take?

Two to four in most real uses. Round one surfaces the spread and the reasoning; rounds two and three do most of the converging. Stop when the spread stops narrowing, not when it hits zero: a stable, honest range beats a forced consensus, and a stubborn outlier with strong reasoning is information.

Is the Delphi method accurate for demand forecasting?

More accurate than the unstructured alternative it replaces (open discussion dominated by rank and anchoring), which is its real benchmark. Studies of structured group judgment consistently favor it in low-data settings. It cannot outperform good data where data exists, which is why it lives at the launches and shifts where data doesn't.

What's the difference between Delphi and a regular expert panel?

A regular panel talks first and estimates under the influence: of the loudest voice, the first number spoken, and the boss's face. Delphi inverts the order: private estimates first, anonymous summarized feedback, then revision. Same people, same knowledge, but the structure controls the social dynamics instead of being controlled by them.

Can the Delphi method be done remotely?

It's arguably better remotely: anonymity is easier to preserve through forms than around a table, rounds fit between meetings, and a facilitator can run the whole cycle over a shared doc and email in a few days. The method predates the tools that now suit it perfectly.

When should you not use the Delphi method?

When good data exists (quantitative methods will beat panel judgment), when the decision is small or routine (the process costs more than the error), and when speed rules (rounds take days). It's also wasted on questions the panel holds no genuine knowledge about; structure can surface expertise, not create it.