Somewhere in your planning system there is a forecast accuracy report for April. It is worthless, and pretending otherwise is the most dangerous thing a planning team can do this quarter. Demand volatility has not made forecasting harder; it has invalidated the assumption forecasting rests on, which is that next month resembles the months that came before it. The useful conversation is not how to build a better model. It is what to do with the data, what to do instead of a point forecast, and how to stop three months of abnormal history from corrupting the next two years of planning.
Why the models failed, precisely
Statistical forecasting in an enterprise resource planning system is almost always some variant of exponential smoothing, with optional trend and seasonality, or a moving average with a seasonal index. Every one of these methods decomposes history into a level, a trend and a repeating seasonal pattern, and projects them forward. A demand shock is none of those three things. It is a discontinuity, and the model has no category for it, so it distributes the shock across the categories it does have. Collapsed demand gets read as a falling level or a new downward trend. Surged demand gets read as a level shift upward. Either way the model does not conclude that something unprecedented happened; it concludes that the world has changed permanently, and it forecasts the new shape indefinitely. There is a second, quieter failure that costs more money. Safety stock formulas are driven by forecast error, not by forecast level. Three months of enormous error inflates the standard deviation used in every safety stock calculation in the catalogue, so the system will now recommend holding far more inventory across thousands of items, permanently, on the basis of a period that will not repeat. Organisations that let this run unexamined will spend the second half of the year wondering why working capital rose while sales fell.
The data decision nobody wants to make
The single most consequential planning choice available right now is how to treat February through May in your demand history. There are three options and only one of them is defensible. Deleting the period is tempting and wrong; you lose the record and the audit trail, and next year nobody can explain the gap. Leaving it untouched is the default, which means the model treats a pandemic as ordinary demand behaviour. This is what most systems are currently doing. The right answer is to mark it. Most planning tools support outlier flagging, demand exceptions, event calendars or manual history adjustment; the mechanism matters less than the discipline. Record actual demand as it happened for reporting and financial purposes, and hold a separate cleansed demand series for forecasting, with the adjustments documented and dated. Where the tool cannot do this, an offline adjusted history maintained by the planning team is better than nothing. And make the decision per segment rather than globally, because the shock was not uniform.
Three portfolios, not one
Aggregate demand statistics are hiding three entirely different situations. The collapsed group: anything tied to travel, hospitality, events, discretionary retail, construction sites that stopped, and business-to-business customers who suspended operations. History is intact but irrelevant until the customer base restarts, and the planning question is about restart timing, not demand shape. The surged group: staples, cleaning and hygiene products, packaged food, home and kitchen goods, computing and connectivity equipment, pharmacy lines, delivery-related packaging. Here the risk is the opposite and larger. The surge is partly pantry loading, which means demand has been borrowed from future months, and a model fed this history will order aggressively into a trough. The steady group: maintenance items, regulated consumables, contracted volumes, anything with demand driven by an installed base rather than by consumer behaviour. These items barely moved and should not be touched. Blanket forecast interventions damage them. Segment first, then decide treatment. A planner who spends a week doing this will outperform any model change.
| Segment | Question to investigate | Planning treatment in the article |
|---|---|---|
| Collapsed | When will customers restart? | Review restart assumptions rather than extrapolating the trough. |
| Surged | Is demand borrowed or structural? | Review commitments and keep adjustments documented. |
| Steady | Has the demand driver changed? | Avoid blanket overrides where the historic model still fits. |
Qualitative summary of this article's source text, not a measured outcome or performance estimate.
Scenarios are decisions, not forecasts
The replacement for a point forecast is not a better point forecast. It is a small number of scenarios, each with a leading indicator that tells you which one you are in and a pre-agreed action that follows. Three is the right number: a slow restart, a staged recovery, and a renewed restriction period. For each, write down the demand assumption by segment, the inventory position it implies, the supply commitments you would make or cancel, and the cash consequence. Then identify the indicator you will actually watch, ideally something outside your own sales data, and the date by which you must decide. The output that matters is not the spreadsheet. It is that when the indicator moves, the organisation already knows what it decided to do, and nobody has to convene a committee to think about it from scratch. Every scenario exercise that produces a document but no pre-agreed trigger has failed.
Practical Guidance for Scenario Planning Workshop
- Flag the shock period in demand history rather than deleting or ignoring it. Keep actuals for reporting, hold a cleansed series for forecasting, document every adjustment with a date and a reason.
- Audit safety stock and reorder parameters for error inflation. Forecast error from this period is silently raising recommended stock across the whole catalogue. Freeze automatic recalculation until you have reviewed it.
- Split the portfolio into collapsed, surged and steady before changing anything. Different segments need different treatment, and blanket interventions damage the items that were fine.
- Treat surges as borrowed demand until proven otherwise. Pantry loading and precautionary buying create a trough behind them. Do not commit long-lead supply on the strength of eight abnormal weeks.
- Build three scenarios with named leading indicators and pre-agreed actions. Slow restart, staged recovery, renewed restriction. Each with a trigger, a decision date and an owner.
- Shorten the planning cycle only where it earns its cost. Weekly review for the volatile segment and top revenue items; leave the long tail monthly. Weekly planning across everything exhausts the team within a month.
- Plan supply variability, not just demand variability. Lead times are less reliable than forecasts right now. Differentiate service levels deliberately by product and customer rather than accepting a uniform target you cannot meet.
- Write the assumptions on the face of the plan. Every number circulated should carry the scenario it assumes. Plans that travel without their assumptions get quoted as facts in board papers.
The Regional Angle
Four regional factors make this harder here than the global planning literature suggests, and none of them are in the model. The first is that the shock landed precisely on top of the largest seasonal event in the regional calendar. Ramadan and the Eid period normally produce the year's most pronounced and most reliable demand pattern: a grocery and gifting surge, hospitality and iftar volumes, extended retail hours, a distinctive category mix. This year that pattern ran under restrictions, with gatherings curtailed, hotel and majlis catering largely absent, and consumption shifted decisively into households and delivery. Any seasonal index derived from this year's Ramadan will be wrong, and because most systems weight recent seasons more heavily, it will be wrong in a way that compounds next year. Freeze your Ramadan seasonal factors at their prior values and note why. The second is population. Consumer demand in the Gulf rests on an expatriate base that moves with employment, and job losses, repatriation and delayed family returns change the size of that base rather than its behaviour. No demand model can see this, because it is not a pattern in sales history; it is a change in the number of people. Planners serving consumer categories should be asking human resources and their landlords and their delivery partners what they are seeing, and treating population as an explicit scenario variable. The third is that the summer trough may invert. The regional year has a dependable structure in which a substantial share of residents travel during the summer months, demand softens, and stock plans are built around it. With travel restricted and quarantine requirements in place at both ends of most journeys, a meaningful portion of that population will remain in market through the summer. For grocery, home, electronics and delivery categories the usual trough may not arrive; for airport retail, travel services and the hospitality supply chain, the usual recovery will not either. Do not apply last year's summer curve. The fourth applies to anyone re-exporting or distributing regionally. Your customers sit in markets with completely different restriction timetables, currency positions and government support, and their restarts will not be synchronised. A single regional forecast is now an average of divergent situations. Forecast by destination market, even coarsely, because a distributor whose plan assumes uniform recovery across the Gulf, the Levant, East Africa and the subcontinent is planning for a country that does not exist.
The objection worth taking seriously
The first serious objection is that adjusting history is how planners lie to themselves. If you flag every inconvenient month as an outlier you eventually build a model that describes a world you prefer. Worse, this period may not be an outlier at all: some of the shifts in channel mix, home consumption and remote purchasing look structural, and a planning team that cleanses them out of the data will be permanently behind the market it actually serves. The case for cleansing rests on an assumption about the future that is no more provable than the model's. The second objection is about capacity. Scenario planning is comfortable advice for organisations with a planning function. A mid-sized distributor with two planners, a demand planning module nobody was trained on, and a sales team that submits its forecast late does not have the bandwidth to maintain three scenarios weekly. Told to do so, it will produce worse plans than it would have by keeping things simple and checking stock cover twice a week. Both land. The honest response to the first is to separate the reporting series from the forecasting series and to review the flags quarterly rather than treating them as permanent, so that if the shift turns out to be structural you adopt it deliberately instead of discovering it in a variance report. The response to the second is to scale the method to the team: for a small planning function, two scenarios and a fortnightly review of the top fifty items by revenue delivers most of the benefit, and the single highest-value action remains freezing automatic safety stock recalculation. Sophistication is optional. Not letting an abnormal quarter silently reprogramme your inventory policy is not.
Common Questions
Should we stop using statistical forecasting entirely?
No. Keep it for the steady segment where it still works, and override it where the assumption of continuity has failed. Turning it off across the board replaces a flawed model with nothing.
How long should we keep the shock period flagged?
Review quarterly. If the change persists for three or four quarters, it is a new level rather than an outlier and should be adopted into the baseline explicitly.
What indicator should we watch?
Something outside your own sales data and available faster: mobility and footfall measures, order intake by channel, customer restart confirmations, port and freight activity, and in construction, site reopening notices.
What should we expect over the next twelve months?
Expect the forecast accuracy conversation to be replaced by conversations about bias and stock cover, which is an improvement. Expect the abnormal-history problem to resurface when next year's budgets are built off this year's actuals, and to be argued about in November. Expect vendors to pitch demand sensing and external data signals hard, and expect most of that value to be unreachable until the internal data is segmented and clean. And expect supply lead-time variability, not demand error, to be the constraint that decides who serves their customers well.
Scenario Planning Workshop — we segment the portfolio, protect your history from an abnormal quarter, and leave you with three scenarios that each have a trigger and a decision already made.