What a demand plan actually is
A demand plan is not a single number pulled from a spreadsheet formula — it is an organisational output. It is what a business agrees, after combining statistics with commercial knowledge, it expects customers to buy over a coming horizon, broken down by the time periods, products and locations that the rest of the supply chain needs in order to act. Procurement uses it to decide order quantities and timing. Production or import planners use it to size purchase orders and container bookings. Finance uses it to forecast revenue. Because so many downstream decisions hang off it, the demand plan has to be more disciplined than "what does the sales team think will happen next month."
The demand planning framework, in its simplest form, is a repeating cycle: gather and clean historical data, generate a statistical baseline forecast from that data, layer on judgemental knowledge that the statistics cannot see (a new customer contract, a competitor stock-out, a planned promotion), reconcile any disagreements into one agreed number, and then measure how accurate that number turned out to be once actuals come in — feeding the lesson back into the next cycle. Done well, this is a continuous discipline, not a once-a-year budgeting exercise.
Three structuring dimensions: time, product, geography
Before choosing a forecasting technique, a demand planner has to decide the level of granularity along three dimensions, because forecasting at the wrong level either drowns the plan in noise or hides information the business needs.
Time. Should demand be bucketed daily, weekly or monthly? Finer buckets (daily) are noisier and harder to forecast accurately but are what a warehouse replenishment system actually needs to schedule labour and trucks. Coarser buckets (monthly) smooth out random day-to-day noise and are usually accurate enough for procurement and container-booking decisions made months in advance. Most mature demand planning processes forecast at a coarser bucket for the far-out horizon and progressively refine to a finer bucket as the period gets closer — a rolling zoom rather than one fixed resolution.
Product. Forecasting every individual SKU (Stock Keeping Unit) — every size, colour and pack-size variant — produces a huge number of low-volume, erratic series that are genuinely hard to forecast well. Forecasting at the product-family or category level (aggregating related SKUs) produces a smoother, more statistically forecastable series, because random ups and downs in individual SKUs cancel out when combined. The common approach is to forecast total family demand statistically, where the pattern is stable, and then split that family total down to individual SKUs using historical mix ratios — a technique sometimes called top-down forecasting with disaggregation.
Geography. Should the forecast run at national level or per distribution centre / region? A national total is more stable and easier to forecast; a per-DC forecast is what is actually needed to decide how much stock to hold where. As with product, a common pattern is to forecast nationally where the statistics are most reliable, then allocate down to each location using historical or planned regional splits.
The general principle behind all three dimensions is the same: statistical forecasting works better on aggregated, smoother data, but operational decisions need disaggregated, granular data. Good demand planning processes deliberately forecast at the aggregate level and cascade down, rather than trying to statistically forecast the noisiest, most granular level directly.
The demand planning process: cleanse, baseline, override
The process typically runs in three stages. First, data cleansing: raw sales history is not the same as true underlying demand. A stock-out that suppressed sales for two weeks, a one-off bulk order from a single customer, or a data-capture error will all distort a naive read of "what we sold." Cleansing means identifying and adjusting for these anomalies so the statistical engine trains on a realistic demand signal rather than a supply-constrained or noisy one.
Second, a statistical baseline is generated from the cleansed history using one or more of the techniques described below. This baseline is a purely mathematical extrapolation of past patterns — it has no knowledge of a new contract signed last week or a competitor exiting the market.
Third, judgemental override and consensus: sales, marketing and customer-facing teams review the statistical baseline against what they know is happening in the market, and adjust it where they have information the statistics cannot capture. This is deliberately a conversation, not a unilateral override — a well-run process requires the adjuster to document the reason for the change, and reconciles conflicting views (sales wants to plan high, finance wants to plan conservatively) into one single agreed number that the whole business commits to and plans against. This reconciliation step is the direct input into the Sales and Operations Planning (S&OP) process, where the demand plan is balanced against supply capability at a cross-functional level.
The main statistical techniques, in plain English
None of the techniques below require heavy mathematics to understand conceptually — each is really just a different way of answering the question "what should I assume tomorrow looks like, based on what I've seen recently?"
| Technique | What it does, intuitively | Best suited to |
|---|---|---|
| Moving average | Averages demand over the last few periods (e.g. the last 3 months) and uses that average as next period's forecast — each new period drops the oldest one and adds the newest | Stable, flat demand with no trend or seasonality; simple, easy to explain to non-specialists |
| Exponential smoothing | Also averages the past, but weights recent periods more heavily than older ones, so the forecast reacts faster to genuine shifts in demand while still smoothing out one-off noise | Demand that drifts gradually over time; extended versions handle trend and seasonality by smoothing each pattern separately |
| Regression-based forecasting | Finds the mathematical relationship between demand and one or more explanatory factors (time, price, marketing spend, temperature) and projects that relationship forward | Demand that is driven by an identifiable external factor, or a clear long-term trend line |
In practice, no single technique wins for every product. Mature demand planning software runs several techniques against the historical data for each product family, tests how well each would have predicted recent actuals, and automatically selects (or blends) the best performer — a process usually called best-fit or champion-challenger forecasting. The planner's job shifts from doing the arithmetic to sense-checking the output and deciding where judgement should override it.
Measuring accuracy — and forecast value added
A forecast that is never measured against what actually happened is just an opinion. Forecast accuracy is usually expressed as some form of percentage error — how far off, on average, the forecast was from actual demand, expressed as a percentage so it can be compared across products of very different sizes. The exact formula matters less than the discipline: track it consistently, by product family and by planner, and review it regularly so systematic problems surface early rather than being discovered when a warehouse runs out of stock or a container arrives with no orders to fill it.
A related and often more revealing question is forecast value added (FVA): does each layer of effort in the process actually improve the forecast, or does it just add cost and complexity without improving the outcome? The test is simple in concept — compare the accuracy of the final, human-adjusted forecast against the accuracy of a naive forecast (for example, simply "next period will equal this period," with no statistics or judgement applied at all). If the sophisticated, heavily-worked forecast is no more accurate than the naive one, the process is adding effort without adding value, and it is worth asking whether the statistical layer, the judgemental overrides, or both, are actually helping. FVA analysis has a habit of embarrassing forecasting processes that everyone assumed were working well.
Biased forecasts: the silent accuracy killer
Accuracy metrics capture how far off a forecast was, but they can hide a more damaging problem: bias, meaning the forecast is systematically wrong in the same direction, period after period. A forecast that is sometimes 10% too high and sometimes 10% too low is inaccurate but unbiased — the errors cancel out over time and the average lands close to right. A forecast that is consistently too high, month after month, is biased, and every downstream plan built on it inherits a permanent structural error.
Bias often has a human cause rather than a statistical one. A sales team incentivised on hitting targets may consistently inflate the forecast to justify more stock allocation, or consistently sandbag it (forecast low) to make the eventual result look like an over-achievement against target. Because standard accuracy metrics measure magnitude of error, not direction, a biased forecaster can look statistically "reasonably accurate" on average while quietly wrong-footing procurement every single cycle. Tracking bias separately from accuracy — and being willing to ask uncomfortable questions about who benefits when a forecast is consistently high or low — is a core part of a mature demand planning process, and part of the same discipline this site covers in the bullwhip effect, where distorted signals amplify as they move upstream.
Special cases worth knowing about
A handful of demand patterns break the standard techniques above and deserve their own treatment, even if a deep dive is beyond this article's scope:
- New product phase-in and phase-out. A brand-new SKU has no history to extrapolate from, so its early forecast has to be built by analogy to a similar existing product, or from market sizing, rather than from its own sales data — and confidence should be flagged as low until real sales history accumulates.
- Promotions and price-based planning. A promotional spike is not "normal" demand and, if left in the history unadjusted, will distort the baseline for future non-promotional periods. Promotional demand is usually planned and tracked as a separate uplift layered on top of the baseline forecast, not folded into it.
- Sporadic or intermittent demand. Some products (slow-moving spares, niche variants) sell in occasional, lumpy bursts rather than a steady stream. Standard smoothing techniques tend to under- or over-react badly to this pattern, and specialised intermittent-demand methods (or simply managing these lines by reorder point rather than statistical forecast) are usually a better fit.
The South African angle: festive season, Black Friday, and the 8-10 week overseas lead time
South African demand planners face a particularly sharp version of the seasonal-spike problem. The festive season (roughly early December through mid-January) and Black Friday represent a hugely disproportionate share of annual retail demand compressed into a few weeks, and a forecast that simply extrapolates the rest of the year will badly under-call these peaks — see our festive-season import planning guide for how that forecast then flows into an actual ordering timeline. Because most imported stock — clothing, electronics, homeware, toys — is ordered from Asian or European suppliers with an 8-10 week combined production, shipping and customs-clearance lead time, the order that determines whether shelves are full for Black Friday has to be placed in September, based on a forecast made even earlier, well before any real-time read on how the season is actually trending.
This lead time also has to be planned around South Africa's public holiday calendar and school term dates, both of which shift underlying consumer demand patterns (school-uniform and stationery demand around January and July school-term starts; a general retail lull around the Easter and Heritage Day long weekends followed by a pickup). A demand planner ordering stock to arrive in time for a specific date has to work backwards not only from the calendar date itself but from the realistic port and customs-clearance time at the South African end — an importer who has learned the hard way not to assume container availability equals shelf availability. A forecast that ignores this compounding lead time, however statistically elegant, will still leave stock arriving after the season it was meant for.
Frequently asked questions
What is the difference between a demand plan and a sales forecast?
A sales forecast is often a single number owned by the sales or finance function, sometimes influenced by target-setting incentives. A demand plan is the supply-chain-facing output of a structured process that reconciles a statistical baseline with commercial judgement into one agreed number the whole business plans against — deliberately separated from sales targets so it reflects expected reality rather than aspiration.
Which forecasting technique is "best"?
There is no single best technique — it depends on the demand pattern of each product. Mature processes test several techniques against recent history for each product family and let the best-performing one win, rather than mandating one method for the whole portfolio.
Why does forecast value added matter if we already track accuracy?
Accuracy alone tells you how wrong the final forecast was, not whether the effort spent getting there was worthwhile. Forecast value added compares each stage against a naive baseline, exposing cases where statistical modelling or manual adjustment is adding cost and complexity without actually improving the forecast.
How can a business tell if its forecasts are biased rather than just inaccurate?
Track the direction of the error, not just its size. If a product or category consistently forecasts high or consistently forecasts low over many periods, rather than the errors landing roughly evenly above and below actuals, that is a sign of bias — often traceable to an incentive structure that rewards inflating or sandbagging the number.
Why does an SA importer need to forecast so far ahead of a seasonal peak?
Because of the combined production, shipping and customs-clearance lead time on imported stock, often 8-10 weeks. An order meant to be on shelves for Black Friday or the festive season has to be placed months earlier, based on a forecast made before any real-time signal on how that season is actually trending — which is exactly why the statistical baseline plus judgemental process described above matters more, not less, the further ahead the forecast has to reach.