AI demand forecasting for distributors: a practical guide

The short answer

AI demand forecasting uses your past sales and other signals to predict how much of each item you will sell, then turns that into a suggested purchase order that respects supplier lead times and minimum order quantities. It helps most when you stock hundreds or thousands of items, demand is uneven and a single min-max rule cannot fit every product. Buyers stay in charge: they see the suggestion, the reason and the numbers, and they approve or change it.

Our demand forecasting and purchase orders page describes the system we build. This article explains when you need it and how to judge it.

When spreadsheets and min-max levels stop working

Min-max reorder levels are a good start. You set a minimum, you order up to a maximum, and the rule is easy to explain. They break down in predictable ways:

• Seasonality. Demand for fans, pumps, spare parts or school supplies swings by month, but the min-max number is fixed.
• Trends. A product that is growing or fading keeps getting the old level.
• Long and variable lead times. Imports from a supplier that sometimes takes 3 weeks and sometimes 7 need a different buffer from a local one.
• Too many items. No buyer can review 5,000 SKUs by hand every week, so most items get stale settings.
• Promotions and one-off orders. A big tender or a festival offer distorts the history, and the spreadsheet cannot tell.

The symptoms are familiar: stockouts on fast movers, cash tied up in slow stock, and emergency orders at higher prices.

The data you need

Data	Why it matters	Common source
Sales history by item and date	The base of every forecast. Two or more years is better, to see seasons	Tally, Zoho, ERP
Stock on hand and on order	You order the gap, not the demand	Inventory module
Supplier lead times	Sets how far ahead you must look	Past PO and receipt dates
Minimum order quantity and pack size	Turns a need into an orderable quantity	Supplier price lists
Storage limits and budget	Caps how much you can hold or spend	Item master, finance
Promotions, price changes, stockout days	Explains unusual sales so they are not read as true demand	Sales team, price list history

One caution: sales are not the same as demand. If an item was out of stock for two weeks, zero sales means lost demand, not no demand. Good systems mark stockout days so the model does not learn the wrong lesson.

Forecasting approaches in plain words

• Moving average. Predict the average of the last few weeks. Simple, stable and surprisingly hard to beat for steady items.
• Seasonal methods. Adjust the average using patterns that repeat each year, such as monsoon or festival peaks.
• Statistical time-series models. Learn trend and seasonality for each item from its own history.
• Machine learning models. Use many signals together, such as price, promotions, weather, calendar and related items, and learn across products. These help most for large ranges, new items with little history and promotion-heavy categories.
• Intermittent demand methods. Slow movers that sell once in a while need their own approach, because averages mislead.

No single method wins everywhere. A sensible system tests several per item class against your history and picks what works, instead of forcing one model on every SKU.

From forecast to purchase-order recommendation

The forecast is the middle of the process, not the end. The full chain looks like this:

1. Forecast demand for each item over the supplier's lead time plus the review period.
2. Add a safety buffer based on how variable demand and lead time have been.
3. Subtract stock on hand and open purchase orders.
4. Round up to the supplier's minimum order quantity and pack size, and cap by storage space and budget.
5. Group the lines by supplier into a draft PO, with the reason for each line.

A simple version of the maths: reorder point = demand during lead time + safety stock. For example, if an item sells 105 units a week, the lead time is 3 weeks and weekly demand has a standard deviation of 12 units, then demand during lead time is 315. A common safety stock formula is service factor x standard deviation x square root of lead time in weeks. With a factor of 1.65 (roughly a 95% service level) that is 1.65 x 12 x 1.73, about 34. The reorder point is about 349 units.

How buyers stay in control

Trust is earned, not assumed. Good practice is to show each recommendation with the forecast, the recent sales it is based on, current stock and the reason for the quantity. Buyers can edit any line, reject it with a reason and add items the system missed. Their changes are logged and fed back, so the system learns where it was wrong. Start in "advisory" mode, where the system only suggests and buyers place orders as before. Move to approval-based drafts only after the numbers have earned it. This is the human-in-the-loop pattern applied to purchasing.

How to measure accuracy

Track four things, by item class and not only in total:

• MAPE (mean absolute percentage error): the average size of the miss, as a percentage of actual sales.
• Bias: whether you consistently over- or under-forecast. Bias hurts more than random error because it is systematic.
• Stockout rate: days or orders where an item was not available.
• Excess stock: value or days of stock above what demand justifies.

A worked example (hypothetical)

This uses made-up numbers for one item over four weeks.

Week	Forecast	Actual	Error (F - A)	Absolute % error
1	110	100	+10	10.0%
2	100	120	-20	16.7%
3	100	90	+10	11.1%
4	120	110	+10	9.1%

Formula: MAPE = average of |forecast - actual| / actual. Here that is (10.0 + 16.7 + 11.1 + 9.1) / 4, about 11.7%. Bias = total error / total actual = +10 / 420, about +2.4%, a mild tendency to over-forecast. If this item costs Rs 400 per unit, carrying 10 extra units ties up Rs 4,000. Across 2,000 items with similar bias the idle stock becomes serious, which is why bias deserves its own line on the dashboard. Compare any AI model against a simple moving average on the same weeks. If it does not beat the baseline, it has not earned its place.

Common mistakes

• Treating stockout days as zero demand. The model then under-orders the very items you ran out of.
• One accuracy number for everything. Fast, slow and new items behave differently, so measure them separately.
• Dirty item masters. Duplicate codes and renamed items break the history. Clean them first.
• Ignoring supplier reliability. A perfect forecast fails if the supplier is late. Track real lead times.
• No dashboard. If buyers cannot see accuracy and exceptions, they will not trust it. A BI dashboard built for your team fixes this.

How to get started

1. Export 12 to 24 months of sales by item and date, with current stock and supplier lead times.
2. Pick one supplier group or category, ideally 100 to 300 items, as the first scope.
3. Record how you decide quantities today, and how often you stock out or hold excess.
4. Send these through our contact page.

We then back-test forecasts on your history and show the MAPE, bias and stock impact against your current method before you commit. If you run a store network, see AI for retail and logistics for related use cases, and the custom software and BI dashboards service for the buyer screens.