DEMDemandStop 12 of 12
- Forecasting fundamentals
- Demand patterns & classification
- Exponential smoothing & ETS
- Intermittent demand: Croston, SBA, TSB
- ML & global forecasting models
- Forecast accuracy & bias
- Probabilistic forecasting
- Hierarchical forecasting
- Promotions, cannibalization & halo
- New products & phase-outs
- Demand sensing & shaping
- Consensus forecasting & FVA
Consensus forecasting & FVA
Letting sales, marketing and planner judgment into the forecast only where it actually makes the forecast better.
Read atChange any time
Most companies don’t ship the statistical forecast untouched. Sales adds customer intelligence, marketing adds promotions, planners make overrides, and a monthly meeting agrees one consensus forecast. Human input can add real information that the data doesn’t have, such as a big new customer or a competitor’s stockout. It can also add bias: sales teams may lowball to beat their targets, or inflate numbers to secure stock.
Forecast value add (FVA) checks whether each step earns its place. Lay the process out as a ladder and measure the accuracy at every rung:
- a naive forecast (“same as last period”), which is free,
- the statistical forecast,
- the planner’s adjusted forecast,
- the final consensus forecast.
FVA for a step is its accuracy minus the accuracy of the step before. If the statistical model has 30% error and planner overrides take it to 34%, those overrides have negative value: they cost time and made things worse. Studies of company forecasting processes have found that many small adjustments add little, while fewer, larger adjustments based on real information do better.
FVA isn’t about removing people. It’s about directing their effort to the items and events where they know something the model doesn’t, and measuring bias as well as accuracy for every contributor.
Words to use
- Forecast value add (FVA)
- Change in accuracy from one process step to the next, compared with a naive baseline.
- Consensus forecast
- The single demand plan agreed across sales, marketing, finance and supply.
- Override
- A manual change to the statistical forecast; worth keeping only if it improves accuracy.
The FVA ladder
Measure every step of the forecasting process on the same items, periods and lag, with the same error measure:
| Step | WMAPE | Bias | Value added vs previous step |
|---|---|---|---|
| Naive forecast | 40% | 0% | (baseline) |
| Statistical forecast | 30% | +1% | +10 points |
| Planner overrides | 33% | +6% | −3 points |
| Final consensus | 31% | +4% | +2 points (still −1 against statistical) |
Reading it:
- The statistical model earns its keep: 10 points better than doing nothing clever.
- Planner overrides make the forecast worse and add a 6% upward bias, costing accuracy and building excess stock.
- The consensus meeting recovers some ground but still ends below the plain statistical forecast.
What to do with the result
- Break FVA down by item segment, contributor and type of adjustment. Overrides may help on promoted items and hurt on stable ones.
- Require a reason code for every override (promotion, new customer, competitor event and so on), so value can be measured by reason.
- Set thresholds: let small, stable items run untouched, and focus judgment on large or uncertain ones.
- Share the findings every month; people improve when they see their own numbers.
The monthly demand review
A typical cycle: close the month’s data, run the statistical forecast, review exceptions (large errors, bias, big changes), collect sales and marketing input with reasons, hold the consensus meeting to agree one unconstrained demand plan, and hand it to supply planning and S&OP. FVA measurement runs across the whole cycle.
What research says about adjustments
A large study of more than 60,000 forecasts at four supply-chain companies (Fildes, Goodwin, Lawrence and Nikolopoulos, 2009) found patterns most companies will recognize:
- Large adjustments tended to improve accuracy; small ones mostly added noise.
- Downward adjustments were more often helpful; upward ones were frequently over-optimistic.
- Adjustments worked best when based on specific information the model couldn’t have, such as a known promotion or a lost customer.
So judgment is valuable, but it should be rare, large and justified.
Where bias comes from
- Incentives: sales teams measured against quota may forecast low to beat it, or high to secure scarce stock.
- Optimism: new products, favourite customers and planned campaigns attract hopeful numbers.
- Anchoring and recency: last month’s number or the latest big order weighs too heavily.
- Hierarchy: in a meeting, the most senior opinion can override the evidence.
Designing a process that adds value
- Separate forecast, target and plan. The demand review produces an unconstrained forecast; targets and constrained supply plans are separate numbers, discussed openly.
- Structure the input. Ask for specific information, such as customer plans, events and price moves, rather than “your number”, and translate it into adjustments.
- Make value visible. Publish FVA and bias by contributor regularly, without blame, so people learn where they help.
- Use statistics by default. Let the model forecast most items untouched, and spend meeting time on exceptions and big decisions.
- Record decisions. Note what was changed and why, so later reviews can check whether the reasoning held.
Is a meeting needed at all?
For stable items, often not: an automated forecast with exception alerts can be better and cheaper. The consensus meeting earns its place on the few items and events where people know something material, and as the moment the business commits to one demand number it will plan against.
Interview kit
Practise it out loud
Answer each question yourself before opening the model answer.
Human input should only stay in the forecast where it adds value. FVA measures accuracy at each step, from naive to statistical to planner override to consensus, and compares each with the step before. Steps with negative FVA, often frequent small overrides, should stop, while targeted adjustments based on real information, such as a new customer or a promotion, are kept. I’d track bias by contributor too, since incentives often skew sales input.
Planner overrides make the forecast worse on average. What do you do?
Measure FVA by planner and by item group to see where overrides help and where they hurt. Restrict overrides to cases with new information, require a reason code, and set a minimum size so small tweaks stop. Track bias too, because systematic over- or under-forecasting often comes from incentives.
Why compare against a naive forecast?
Because it costs nothing. If an expensive model or a long meeting can’t beat “same as last period”, the process is destroying value. The naive forecast is the honest baseline for FVA.
Connected stops
Where this idea feeds in, or gets its inputs from.
Sources and further reading
- Gilliland, M. (2010). The Business Forecasting Deal. Wiley.
- Fildes, R., Goodwin, P., Lawrence, M. & Nikolopoulos, K. (2009). Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply-chain planning. International Journal of Forecasting, 25(1), 3–23.
- Fildes, R. & Goodwin, P. (2007). Against your better judgment? How organizations can improve their use of management judgment in forecasting. Interfaces, 37(6), 570–576.