Forecasting demand across 60 stores to cut wastage by a third

Sattva operated 60 stores selling fresh and short-shelf-life products. Store managers ordered on intuition, which produced a persistent trade-off: order conservatively and lose sales to empty shelves, or order generously and write off unsold stock.

Wastage was running at 11.2 percent of fresh category revenue, while availability at 6pm sat at 78 percent.

We built a store-and-product level demand forecast using two years of sales history combined with day of week, local holidays, weather and promotion calendars. Output feeds an ordering recommendation the store manager can accept or override.

We deliberately kept the manager in the loop rather than automating the order outright, both because local knowledge genuinely matters and because trust had to be earned before autonomy was reasonable.

The data assessment took three weeks and produced an uncomfortable finding: point-of-sale data did not distinguish a genuine zero-demand day from a day the product was simply unavailable. Training on that data would have taught the model to under-order exactly the items that sold out. We reconstructed censored demand using stock-out timestamps before any modelling began.

Rollout was staged across eight stores for six weeks, with a matched control group, so the effect could be measured rather than assumed.

Wastage fell from 11.2 percent to 7.4 percent of fresh category revenue, a 34 percent reduction. Availability at 6pm improved from 78 percent to 91 percent, so the gain did not come at the cost of empty shelves.

Manager override rates fell from 44 percent in the first month to 12 percent by month four, which we read as the clearest signal that the forecast had earned its place.