The conversation we have most often with consumer goods leaders in India is the one about the forecast. The factory schedules off it. The trade marketing team plans schemes against it. Finance projects working capital with it. The finance head signs cheques based on it. And yet, when we audit the forecast accuracy in most mid-sized brands, the number is the same: 30 to 40 percent error at SKU-store-week grain, sometimes much worse.
The team has usually accepted this as a fact of life. It is not. A working forecast routinely runs at 12 to 18 percent error in the same business, on the same data, in six to eight weeks of build time. The gap between today and a working forecast is not a multi-year transformation. It is a handful of decisions about how the forecast is constructed and who looks at the output every week.
What the forecast actually is, today
In almost every Indian consumer brand under five hundred crore in revenue, the forecast is one of three things, and often a mix:
- A rolling three-month or six-month average, computed in Excel by the demand planner.
- An annual operating plan number, divided by twelve, with seasonal index applied by hand.
- A sales head's call on what next month "should" be, based on the trade meetings of last week.
The output of all three is a flat or near-flat line. The actual demand is anything but flat. The picture below is the gap, in a typical category.
The forecast on the dotted line is the rule-of-thumb. The forecast on the solid green line is the model-driven one. The actual demand is the black line that they are both trying to predict. The black-and-green gap is what your team should be measuring. The black-and-red gap is what your team is living with.
Why the rule-of-thumb is so wrong
Three structural reasons.
The data is at the wrong grain. The plan is national, the demand is local. A 500 gram pack of a snack moves at a different rhythm in Chennai versus Indore versus modern trade in Mumbai. A national average is the wrong unit for a forecast that drives store-level replenishment.
Promotions are folded in invisibly. The history the forecaster looks at is already mixed: baseline plus scheme volume plus festival shifts. The team plans next month from a history that contained five overlapping schemes. The forecast captures none of them, then the scheme calendar changes.
The team that owns the forecast is not the team that learns from it. Forecast errors do not get reviewed in the planning meeting. They get apologised for in the variance meeting two weeks later, and the next forecast is built without using the lesson.
The five inputs of a forecast that works
A working forecast uses five inputs at minimum. Most teams use only the first one.
Sales history at SKU, store, and week grain. Not monthly, not national. The model needs the same level the decision is made at.
Seasonality captured properly. Festivals, monsoon, year-end, school calendars. Each has a specific effect on different categories. Diwali lifts snacks but not detergent. Monsoon lifts ORS and umbrellas but tanks ice cream. The model has to know.
Promotions, scheme by scheme. The scheme calendar feeds the model. The model knows the scheme is coming, the model knows what the lift was the last three times the scheme ran. The model takes the lift out of history when training and puts it back in when forecasting forward.
Calendar events. Cricket matches, school holidays, election days, long weekends. They move beverage volume, they move out-of-home volume, they move impulse categories.
External signals where they help. Weather, retailer-level mood (from sales-team notes), competitor activity. These do not always help. When they do, they help meaningfully.
A six-week build, end to end
The fastest path is honest about what to do in what order.
Weeks one and two, data pull and join. POS or sell-out data, ERP for sell-in, the promotion calendar, the master data for SKU and store hierarchies. The join is usually the unglamorous bit nobody wanted to do. Without it, nothing else works.
Weeks three and four, model train and validate. Start with a simple time-series model, run it against a hold-out test period, see the error. Add seasonality, add promotion effect, retest. Add calendar events, retest. Do not chase a 90 percent accurate model in weeks three and four. Land 75 percent in week four and improve.
Weeks five and six, put it into the weekly planning meeting. The forecast is read by the demand planner, by the sales head, by finance. The team owns the forecast and edits where they have ground-truth that the model does not.
What tends to go wrong
The model is usually fine. The team usage is where the build either lands or stalls.
The team does not trust the model in week six. Fine. The model is wrong by 18 percent. The rule-of-thumb is wrong by 32 percent. Compute both every week for three months. Trust builds when the team sees the model beat their gut consistently.
The team overrides the model every week. Then the override quality has to be tracked too. If the override beats the model, learn from the override. If the override loses, the team learns to override less often.
The data pipeline breaks every other week. The unglamorous part. The data engineer has to own this. Without a stable pipeline, the forecast becomes the demand planner's Excel by week ten.
What to do this week
Get the forecast accuracy number for the last twelve months at the SKU-store-week grain. Just the number. If you do not have it, that is the diagnosis. Get someone to compute it on whatever data you have. If the number is between 28 and 40 percent, you are typical. If it is over 40 percent, you are leaking margin every week. Either way, the path forward is the same one above.
If you want help running this build with a senior team that has done it many times in Indian consumer goods, talk to us.