Demand Forecasting That Beats the Spreadsheet
The baseline you have to beat is better than you think
An experienced planner with a spreadsheet and category knowledge is a genuinely strong baseline, particularly for slow-moving items and promotional periods. Before starting, backtest the current process against the last two years of actuals and record its error by category. Plenty of forecasting projects have delivered a model that was marginally worse than the planner and nobody noticed until it was live. Knowing the number keeps everyone honest and sets a fair bar.
Data problems will consume most of the timeline
Stockouts are the classic trap: historical sales show zero because there was nothing to sell, not because demand was zero, and a model trained naively learns to under-forecast exactly the items you most want to get right. You need to reconstruct unconstrained demand. Add to that promotional calendars that live in someone's inbox, product hierarchy changes, store openings and closures, and returns booked against the wrong date. Budget more time for this than for modelling; it is not glamorous and it decides the outcome.
Forecast at the level decisions are made
There is no point forecasting daily SKU-store demand if replenishment runs weekly at distribution-centre level. Match the grain to the decision, then aggregate or disaggregate deliberately with a reconciliation step so store, region and national forecasts agree. Mismatched grain is why forecasts that look accurate in a report still fail to improve inventory — nobody can act on them.
Predict intervals, not just points
A single number invites false precision. What a planner actually needs is a range and a service-level implication: order this quantity to hit 95% availability, accept this much expected overstock. Quantile forecasts feed directly into safety-stock calculations and make the trade-off explicit rather than hiding it inside a point estimate. This framing also tends to be what finally wins planners over, because it matches how they already think.
External signals help, selectively
Weather genuinely moves demand in food, beverage, apparel and home categories. Local holidays and school terms matter, and in Gulf markets Ramadan timing shifts seasonality substantially year to year. Macro indicators mostly add noise at the operational horizon. Add external features one at a time and keep only what improves backtested error — most feature wishlists shrink by half once tested.
Ship it into the workflow or it won't be used
The forecast has to arrive inside the planning tool people already open, with the ability to override and a record of who overrode what. Track override rates by planner and category: high override rates on a specific category usually mean the model is missing a real signal that person knows about. That feedback loop is what turns a model into an accepted part of the process rather than a report everyone ignores by month three.
Have a project that needs this kind of thinking?
Let's talk through what you're building — free consultation, no commitment.
Get in Touch