G-XR8P2XJ088

AI Demand Forecasts: Read the Range Around the Estimate

Imagine an AI-assisted planning report saying that next week’s workshop will need 40 materials kits. The organiser books the room, orders supplies and arranges help. Yet the most consequential information may be sitting beside that tidy estimate: how widely could actual demand vary?

A forecast of 40 accompanied by a narrow range presents a different planning problem from the same estimate surrounded by substantial uncertainty. Before turning a forecast into a commitment, give its range a clear meaning. Then decide which commitments can wait, which can change and which need protection.

Ask what the interval is predicting

Start by naming the quantity: kits requested for one particular workshop, for example, with a booking deadline and a defined cancellation rule. Keep requests, confirmed bookings and eventual attendance distinct. Otherwise the forecast and the result may describe different things.

In their forecasting textbook, Rob Hyndman and George Athanasopoulos explain prediction intervals as ranges for future observations with specified probabilities. An 80% prediction interval describes uncertainty about the future quantity under the forecasting method and its assumptions. It does not mean that the single headline estimate has an 80% chance of being exactly right.

Also check whether the report actually supplies a prediction interval. The NIST/SEMATECH statistical handbook distinguishes uncertainty about an estimated average response from uncertainty about a new observation. Predicting that new observation must also allow for its individual variation. A confidence interval for an average therefore answers a different question. NIST also explains that repeated use of the prediction method, under the required conditions, covers new observations at the chosen rate. Individual outcomes can still fall outside.

For the organiser, knowing an average precisely would not remove the uncertainty about next Tuesday’s requests. Ask the report writer to identify the interval type, coverage level and calculation method explicitly.

Two forecasts can share the same headline

Consider two invented reports for the same workshop. These numbers illustrate interpretation; no forecasting model has been fitted or validated for this example.

  • Report A: point estimate of 40 kits; stated 80% prediction interval of 34–46.
  • Report B: point estimate of 40 kits; stated 80% prediction interval of 10–70.

The headline alone cannot tell the organiser which planning situation they face. Report B calls attention to a much broader set of possibilities. It might justify checking whether supplies can arrive in stages or whether a second helper can remain provisionally available. Report A would support a different discussion, subject to evidence that its narrower range is credible.

Neither upper endpoint is an automatic order quantity. Neither lower endpoint promises that demand cannot fall further. The interval should inform a decision alongside the consequences of unused materials, missed requests and inflexible arrangements.

Require evidence for the percentage too. If someone has simply chosen an optimistic number and a pessimistic number, label them as planning assumptions. MaryChuks’ guide to using scenarios and “what if?” thinking offers a useful way to work with such possibilities without assigning unsupported probabilities.

Check past forecasts from the right point in time

A persuasive explanation of a range is only part of the review. Ask how the forecasting process performed when its outcomes were still unknown.

Hyndman and Athanasopoulos describe rolling-origin evaluation: repeatedly forecast a later period using only information available before it, then compare the prediction with the observed outcome. They also explain why evaluation should match the required forecast horizon. A next-day test does not directly establish performance four weeks ahead.

For the workshop business, I would request a dated record of forecasts issued at the same booking lead time as the current decision. Preserve what was known then. A final attendance figure added afterwards must not quietly become an input to its own forecast.

Compare methods on the same workshops and deadlines. Include a clearly described simple baseline, such as a forecast based on recent comparable sessions. If an AI-assisted process changes both the model and the evaluation cases, its claimed improvement needs a closer explanation.

Count misses and examine the width

Once outcomes arrive, record how often they fell inside the stated intervals and how far outside they landed when missed. Treat the percentage printed on the report as something to investigate, rather than evidence that the method has achieved it.

Coverage alone can also flatter an unhelpful report. An enormous interval may include many outcomes while offering little guidance for staffing or supplies. A narrow interval can look decisive while repeatedly excluding what actually happens.

In a March 2007 paper on probabilistic forecast evaluation, Tilmann Gneiting and Adrian Raftery discuss an interval score that accounts for width and penalises observations outside the bounds. Its practical lesson is to review those features together. Simply rewarding a report for containing the answer can encourage ranges too broad to help.

Ask whoever evaluates the forecasts to show the number of cases, the coverage level being assessed and the widths on a consistent scale. A handful of successful workshops should prompt continued observation, not a sweeping claim of reliability.

Turn uncertainty into an operating plan

For the hypothetical organiser, I would separate a forecast record from an action record. The forecast states what the method predicts. The action record explains what the organiser chooses to do about it.

They might confirm reusable supplies first, set a deadline for additional orders and agree when provisional help becomes a firm booking. Each action should have a named owner and a trigger that can actually be observed, such as confirmed requests reaching an agreed threshold before the supplier’s deadline.

That threshold is a management choice. It should reflect the workshop’s constraints and the consequences of getting the decision wrong. Calling it an AI recommendation should not obscure who authorised the commitment or why.

Keep a short forecast record

For each forecast used in a decision, retain:

  • The quantity, unit, target period and time the forecast was issued.
  • The point estimate, interval type, stated coverage and lower and upper bounds.
  • The source data, forecasting method and important assumptions.
  • The action chosen, its owner and its review trigger.
  • The eventual observed outcome and the evaluation results.

This is a proposed working record, not a validated forecasting protocol. Its purpose is to make later review possible. When a workshop finishes, the organiser can inspect both the forecast and the decision made from it, without rewriting either to fit the outcome.

Sources


Discover more from Marychuks.com AI, Psychology, Business & CreativeVerse

Subscribe to get the latest posts sent to your email.

One thought on “AI Demand Forecasts: Read the Range Around the Estimate”

Leave a Reply

Discover more from Marychuks.com AI, Psychology, Business & CreativeVerse

Subscribe now to keep reading and get access to the full archive.

Continue reading