Skip to main content

16. Judgmental Forecasting

Teaching rewrite — the Review process step: how people should (and shouldn’t) adjust the baseline. Original wording; illustrative numbers.

Introduction​

After the model produces a baseline, planners, sales, and a consensus meeting may adjust it. These human adjustments are judgmental forecasts. Done well, they add information no model has; done badly, they inject bias, politics, and wasted hours.

FVA (lesson 12) is the foundation — it shows whether people add value. This lesson explains when to adjust, why people go wrong, and how to run group forecasts that work.

This lesson covers:

  • When judgmental adjustments make sense
  • Three sources of bias: cognitive, intentional (misaligned incentives), and biased processes
  • The wisdom of the crowds and its three pillars
  • Assumption-based discussions: vote first, debate after

16.1 When to adjust the baseline​

Rule: adjust only when you have information the model doesn’t.

StatementAdjust?Why
“Our main customer expects lower sales than usual.”✅The model doesn’t talk to customers
“I’ll forecast this new product manually.”✅Models usually can’t forecast launches
“We raise prices next month.”✅ if the model ignores price; ❌ if price is already a featureDon’t double-count
“I think the trend should be higher.”❌ (without new info)Models detect trends; if one is missed, fix the model
“Seasonality looks wrong.”❌ per itemFix the model or engine, not every item by hand
“A big market change is coming.”✅Unknown to the model
“Last month had a special event; I’ll edit history.”❌ edit; ✅ tag the eventLet the model handle tagged events

Judgment is also right when history is thin or the world is changing fast (new competitors, regulation, pandemic, war). These aren’t absolute rules — but they focus effort where people are most likely to add value (efficacy) and least likely to waste time (efficiency).

Pro tip — let adjustments teach you

Patterns in adjustments reveal model gaps. If planners beat the model without extra information, the model isn’t good enough. If they spend time on promotions, add promotions as a driver. If they fix seasonal items, the engine is weak on seasonality.

16.2 Judgmental biases​

Remember what a forecast is — the best unbiased estimate of future unconstrained demand — and what it isn’t: a budget, a sales target, or a supply plan (lesson 2).

flowchart TD
  J[Judgmental biases] --> C[Cognitive biases]
  J --> I[Intentional biases<br/>misaligned incentives]
  J --> P[Biased process]

Cognitive biases (unintentional)​

BiasWhat happensForecasting example
AnchoringThe first number heard pulls estimates toward it — even an irrelevant one (classic experiments used a spun wheel)A manager opens with “we’ll sell 20% more than last quarter”
ConfirmationWe seek and favour evidence for what we already believe or wantMarketing over-estimates its campaign’s uplift, surrounded by like-minded colleagues
ApopheniaSeeing patterns in noise; explaining every wiggleOver-confidence from “patterns” in random history; reports explaining every deviation
Hindsight“I knew it all along”Past shocks feel predictable after the fact

Better than asking “what did we miss?” or “who was right?”: ask “how do we make sure we capture this information next time?”

Intentional biases — misaligned incentives​

Ask three colleagues for next month’s forecast:

PersonNumberHidden motiveName
Sales rep500Bonus for beating the forecast → wants it lowSandbagging
Customer service2,000KPI is fill rate → wants plenty of stockHedging
Finance1,250Must stay on budgetEnforcing

Other examples: sales raising a forecast to match the budget during a slump (enforcing); sales over-forecasting for opportunistic stock (hedging); regional planners inflating forecasts to grab scarce central stock (shortage gaming → bullwhip, lesson 4). People who benefit from an optimistic or pessimistic forecast tend to produce one — especially when no one tracks their accuracy.

Biased process​

ProblemExample
One-sided data or assumptionsCounting store openings but not closures, launches but not discontinuations, our marketing but not competitors’. Most common: using sales instead of demand (ignoring lost sales)
Selective justificationManagement demands reasons for downward changes but waves through upward ones → systematic over-forecasting (or vice versa)
Imbalance of influenceThe loudest person, a powerful channel, or the final signatory steers the number; executives refuse to sign a declining forecast

Asymmetric costs of over/under-stock belong in service levels and safety stock, not in the forecast (lesson 11).

Checkpoint: when and why people go wrong​

Quiz

Checkpoint — sections 16.1 and 16.2

1 / 4

Which adjustment is most justified?

Pick an answer to see if you are right.

16.3 Group forecasts​

Wisdom of the crowds​

In 1906 Francis Galton analysed about 800 guesses of an ox’s weight at a country fair. The average guess was almost exactly right — better than the winner and the experts. The idea — the average of many independent estimates beats most individuals — was popularised a century later as the wisdom of the crowds.

It works only under conditions:

PillarMeaningIn practice
Objective alignmentEveryone tries to estimate the truth — no hidden games. Shared biases don’t cancelAgree what a forecast is (and isn’t); remove incentives to bias; track FVA
Mindset diversityDifferent backgrounds see different information; biases offsetMix sales, marketing, finance, supply; field and big-picture; quantitative and qualitative; optimists and pessimists
IndependenceEach person forms a view alone first — avoid anchoring and groupthinkPrepare numbers in advance; reveal simultaneously or anonymously (shuffle written estimates)

Worked example: averaging vs independence​

Five people estimate next quarter’s demand (true value 1,000): 900, 950, 1,040, 1,080, 1,100.

  • Individual errors: 100, 50, 40, 80, 100.
  • Average = 5,070 / 5 = 1,014 — error 14, better than every individual.
  • If everyone had first heard a manager say “1,300” and drifted toward it (say +150 each), the average would be 1,164 — anchoring makes errors move together, so they no longer cancel.

16.4 Assumption-based discussions​

The baseline says 1,000. Three colleagues, working independently with aligned goals, say 1,250, 1,300, 1,150. What’s the final forecast?

Most people answer ~1,200 (the average). Now ask why:

PersonNumberAssumptionUplift
Sales1,250We cut the price by 10%+250
Customer service1,300We opened two new stores+300
Finance1,150Our main customer placed a big order+150

Each had different, additive information. The right forecast is 1,000 + 250 + 300 + 150 = 1,700 — not 1,233 (the average of 1,250, 1,300, 1,150).

Vote first, debate after: collect numbers independently, then discuss the assumptions behind them.

Good practices:

  • Discuss drivers and assumptions (launches, marketing, competition, pricing — good and bad news), not just final numbers. Arguing over numbers invites the loudest voice and groupthink.
  • Ask open questions: “How did you account for the promotion?” rather than “Did you include it?”
  • Let the highest-paid person speak last.
  • Check for overlap: if two people cite the same price cut, count its effect once.

Checkpoint: group forecasting​

Quiz

Checkpoint — sections 16.3 and 16.4

1 / 4

Which is NOT one of the three pillars of the wisdom of the crowds here?

Pick an answer to see if you are right.

Rank: judgmental forecasting​

Quiz

Judgmental forecasting practice

1 / 2

Rank these adjustments from MOST to LEAST likely to add value.

OrderDrag the handle to rearrange — top should be first.

  1. 1Lowering a forecast after a key customer announced cuts
  2. 2Raising a trend because it “feels low”
  3. 3Many ±1% tweaks across all items
  4. 4New-product forecast based on launch plans

Drag items into the correct order (or use ↑↓), then check.

Check: judgmental forecasting​

Quiz

Judgmental forecasting

1 / 6

When should planners adjust the baseline?

Pick an answer to see if you are right.

Takeaways​

Remember
  • Adjust the baseline only with information the model lacks; fix model gaps at the source
  • Watch cognitive biases: anchoring, confirmation, apophenia, hindsight
  • Remove intentional biases: sandbagging, hedging, enforcing — by aligning incentives and tracking FVA
  • Fix biased processes: one-sided data, selective justification, imbalance of influence
  • Wisdom of the crowds needs aligned objectives, diverse mindsets, and independent work
  • Vote first, debate after — discuss assumptions, add distinct effects, let senior people speak last

Further Reading​