Every crate of fruit that is thrown away at a dark store or a hotel cold room was, at some point, an ordering decision. Order too little and you lose sales; order too much and the extra stock ages on the shelf until it has to be marked down or discarded. With fresh produce, the window between those two mistakes is only a few days wide.
Demand planning is the discipline of narrowing that window. Over the past few years, machine-learning (ML) forecasting has moved from large Western grocers into Indian retail, and the early results are worth understanding, both for what they promise and for where they still fall short.
Why fresh produce is so hard to forecast
Most packaged goods sit on a shelf for months. A mistake in the forecast simply means a little more stock in the warehouse. Fresh produce gives you no such cushion:
- Short shelf life. Leafy greens, berries and cut herbs may last only a few days even in good cold storage.
- Volatile demand. Weather, festivals, weddings, promotions and even a cricket match change what people buy on a given day.
- Variable supply. Harvest timing, rain and transit delays change what is available and at what quality.
- Thousands of combinations. A network of stores or micro-fulfilment centres (MFCs) multiplies every item by every location by every day.
The scale of the problem is large. The UN Environment Programme's Food Waste Index Report 2024 estimated that about 1.05 billion tonnes of food were wasted in 2022 across households, food service and retail, roughly 19 per cent of food available to consumers. Retail alone accounted for about 131 million tonnes. Separately, the FAO estimates that around 13 per cent of food is lost between harvest and the point of sale.
In India, a 2022 study by NABCONS for the government estimated annual post-harvest losses of about 7.36 million tonnes of fruit and 11.97 million tonnes of vegetables, with loss rates for individual crops such as guava and tomato above 11 per cent. Not all of that is a forecasting problem, but a good share of loss further down the chain comes from stock that arrived when nobody needed it.
What AI demand planning actually does
"AI" here usually means ML models trained on historical sales and other signals to predict how much of each item each location will sell over the next few days. Common model families include gradient-boosted trees (such as XGBoost and LightGBM), random forests and, at the more complex end, neural networks.
The inputs typically include:
- daily point-of-sale or order history per item and location;
- calendar effects such as festivals, holidays and store closures;
- promotions and price changes;
- weather, where it proves useful;
- product hierarchy, so a model can learn from similar items.
The output is not just a single number. Better systems produce a range, so the ordering logic can decide how much risk to take on each item.
An Indian example: More Retail
The clearest published Indian case is More Retail Ltd, which runs hypermarkets and supermarkets across the country. In a case study published on the AWS Machine Learning Blog, More Retail and its analytics partner Ganit described building daily forecasts for about 6,230 store-and-item combinations in fruit and vegetables, looking two days ahead.
According to that write-up:
- forecast accuracy rose from 24 per cent to 76 per cent;
- wastage in fresh produce fell by up to 30 per cent;
- in-stock rates improved from 80 per cent to 90 per cent.
Two design choices in that project are useful for any buyer. First, the team tested around 80 hypotheses about what drives demand, including festivals, weather and promotions, and kept the ones that helped. Second, the system did not just order "the forecast". It generated forecasts at several levels of confidence and chose a higher one for items that had recently gone out of stock, and a lower one for items with high wastage. Order quantities then factored in stock on hand, minimum order quantities and display needs.
These are vendor-published figures from a single retailer, not an independent audit, so treat them as an indication of what is possible rather than a guaranteed result.
Where the models still struggle
A 2026 paper in the journal Oeconomica, using sales data for perishable product families from a supermarket chain, looked specifically at over-forecasting, the error that turns directly into waste. Its findings are a useful reality check:
- ML models (random forest and XGBoost) reduced the size of over-forecast errors compared with a simple seasonal benchmark.
- Promotions and holidays both pushed over-forecasting up, with holidays causing the most instability.
- The model with the best headline accuracy score was not the model with the lowest estimated waste.
That last point matters. A supplier or platform may quote an impressive accuracy percentage, but if the measure averages out errors in both directions, it can hide consistent over-ordering on the items that spoil fastest. For perishables, the question to ask is how the model performs on waste and on high-risk days, not only on average.
Where the waste reduction comes from
ReFED, a US non-profit that analyses food-waste solutions, groups enhanced demand planning with inventory management under what it calls refining product management. The logic is simple: match purchases to real sales as closely as possible, so orders do not create excess supply. In practice, the savings come from a few places:
Lever | What changes |
|---|---|
Item-level forecasts | Slow sellers are ordered in smaller, more frequent drops |
Event awareness | Festivals and promotions are planned for instead of guessed |
Risk-based ordering | Fast-spoiling items are ordered more conservatively |
Shorter lead times | Forecasts only need to be right for one or two days |
Shared data | Suppliers can pack and dispatch to the expected demand, not a fixed standing order |
What this means for trade buyers
You do not need your own data-science team to benefit. Most of the gain comes from better information flowing between buyer and supplier.
Share what you can
- Sell-through, not just orders. A supplier that sees what actually sold, and what was written off, can plan grades and pack sizes far better than one that only sees purchase orders.
- Event calendars. Tell suppliers early about promotions, menu changes, banquets, corporate events or store openings.
- Write-off reasons. "Overripe on arrival" and "did not sell" are different problems with different fixes.
Ask any supplier
- How far ahead can you confirm availability for imported and seasonal lines?
- What is your cut-off time for next-day orders, and can it move closer to dispatch?
- Can you split a delivery into smaller, more frequent drops for short-life items?
- Can you pack to the ripeness stage I need on the day, rather than a single fixed stage?
Measure the right thing
If you use a forecasting tool, track wastage and stock-outs by item alongside headline accuracy. Pay particular attention to festival weeks and promotions, when even good models over-forecast.
The limits of the technology
AI forecasting does not fix a broken cold chain, poor grading or long lead times. A perfect forecast is wasted if produce arrives too ripe, or if the supplier cannot deliver the quantity ordered. Forecasting works best as part of a short, well-run chain: graded stock, packed to specification, held at the right temperature and delivered quickly.
It also needs clean data. Missing days, unrecorded write-offs and items coded differently across locations all weaken a model. The More Retail team, for example, had to handle gaps and outliers in two years of sales history before building its models.
Getting started
For most buyers the practical first steps are modest: keep an item-level record of what sells and what is discarded, share it with your main suppliers, and move fast-spoiling lines to smaller, more frequent orders. That alone narrows the gap between what arrives and what sells, whether or not a machine-learning model sits behind it.
If you would like to discuss order patterns, pack sizes or delivery frequency for your locations, ask ORIGO for today's availability and a quote through the contact page.
Sources
- AWS Machine Learning Blog, "From forecasting demand to ordering: An automated machine learning approach with Amazon Forecast to decrease stock-outs, excess inventory, and costs" (More Retail case study): https://aws.amazon.com/blogs/machine-learning/from-forecasting-demand-to-ordering-an-automated-machine-learning-approach-with-amazon-forecast-to-decrease-stock-outs-excess-inventory-and-costs/
- Albu, A. O., "When Accurate Forecasts Still Produce Waste: Demand Shocks and Over-Forecasting Risk in Food Retail", Oeconomica: https://www.oeconomica.org.ro/abstract/697/When-Accurate-Forecasts-Still-Produce-Waste-Demand-Shocks-and-Over-Forecasting.html
- ReFED, "Following the Roadmap: Refine Product Management": https://refed.org/articles/following-the-roadmap-refine-product-management/
- Rural Voice, report on the UNEP Food Waste Index Report 2024: https://eng.ruralvoice.in/one-fifth-of-all-food-available-to-consumers-eventually-goes-to-waste-a-new-unep-report-reveals
- FreshPlaza, "India reports high guava and tomato losses" (NABCONS post-harvest loss study): https://www.freshplaza.com/asia/article/9756104/india-reports-high-guava-and-tomato-losses/
