A snacks brand in Coimbatore reorders when stock drops below 500 units, a rule set three years ago when the number felt about right. It has never been revisited. During Diwali the brand stocks out for nine days. In February it carries four months of cover on the same SKU. The rule is not wrong so much as frozen, applied to a demand pattern that moves constantly.
Reorder points are usually the highest-leverage number in an inventory system, and usually the least maintained.
The Formula Is Only as Good as the Variance
The textbook reorder point is average daily demand multiplied by lead time, plus safety stock. The first half is easy. The safety stock term is where most implementations go wrong, because they set it as a flat number of days rather than deriving it from actual variability.
Safety stock should reflect two sources of uncertainty: how much demand swings around its average, and how much lead time swings around its average. A supplier with a 21 day average lead time that occasionally takes 40 days needs materially more cover than one that reliably delivers in 25. Service level is the dial. Moving from a 90 percent to a 95 percent service level typically raises safety stock by around 30 percent, and pushing to 99 percent can nearly double it. That tradeoff should be a deliberate commercial decision per product class, not a global default.
Segment Before You Forecast
One model across a whole catalogue produces mediocre results everywhere. Segmentation fixes most of it.
- ABC by revenue contribution. The top 20 percent of SKUs usually drive around 80 percent of revenue and deserve tighter review cycles, weekly rather than monthly.
- XYZ by demand variability. Steady sellers suit simple exponential smoothing. Erratic ones need wider safety buffers and human review, not a more elaborate model.
- Lifecycle stage. New launches have no history worth forecasting, so use analogue SKUs and short review cycles until eight to twelve weeks of real data exists.
- Promotion sensitivity. Products whose sales triple during a sale event must have promotion periods excluded from baseline calculations, or every future forecast inherits a distorted average.
That last point causes more damage than any modelling choice. If a Big Billion Days spike is left in the baseline, the system will over-order for the following quarter and lock up working capital.
Indian Demand Has Its Own Shape
Festive concentration is sharper here than in most markets, with a large share of annual discretionary sales landing in a narrow window between Onam and Diwali. Regional festivals shift by state and by lunar calendar, so a fixed week-of-year seasonality index will drift year to year. Monsoon timing affects categories from footwear to home care. GST rate changes move buying patterns for weeks around the announcement.
Practical response: keep a calendar of demand-affecting events as structured data the forecasting logic reads, rather than as tribal knowledge in a planner’s head. Recompute reorder points on a schedule, monthly for A items and quarterly for the tail. Track forecast error as mean absolute percentage error by segment, and treat any segment consistently above 30 percent as a candidate for manual planning rather than automation.
At Invexa, we favour forecasting systems that stay explainable to the people using them, since a planner who understands why a number moved will trust it, and a planner who does not will override it every time.