Machine Learning for Wholesale Distributors

Distributors sit on unusually good data

Retailers mostly sell to anonymous shoppers. Wholesale distributors sell to known accounts that reorder the same things on recognisable cycles. That makes a distributor's order history some of the most useful data in any sector for machine learning, and it is usually under-used.

Picture a 50-person electrical wholesaler with 14,000 active SKUs, 2,500 trade accounts and 700 order lines a day. Buyers manage reorder points for the top few hundred lines by hand. Sales reps know their biggest customers well and the other 2,000 barely at all. The data to do better is already in the ERP. It just is not being asked the right questions.

Machine learning use cases for wholesalers

1. SKU demand forecasting and safety stock. Setting stock levels from predicted demand and its variability rather than fixed min-max figures.
2. Customer lapse detection. Spotting accounts whose order frequency or basket is shrinking, which usually means a competitor is taking share.
3. Quote conversion prediction. Estimating which quotes will win, so reps chase the right ones and pricing can be tuned.
4. Cross-sell suggestions for reps and web shop. Products similar customers buy that this account buys elsewhere.
5. Dead and slow stock prediction. Flagging lines likely to become obsolete while they can still be returned to the supplier or sold through.
6. Credit risk. Early warning on accounts whose payment behaviour is deteriorating.

Lapse detection is the one we see underrated most. A trade customer rarely announces they are leaving. They just start buying cable from somebody else, then fittings, then everything. Seeing the drift after two missed cycles instead of two quarters makes a real difference to whether a rep call can save the account.

Forecasting thousands of lumpy SKUs

Wholesale demand is lumpy. A contractor orders 400 metres of cable once and nothing for three months. Standard forecasting methods struggle with this intermittent pattern, and many distributors conclude forecasting does not work for them.

It works if you forecast the right things. For fast movers, forecast demand directly. For intermittent lines, forecast how often orders come and how big they tend to be separately, then set stock to cover a target service level. For very large one-off orders, the best signal is often the customer's quote or project pipeline, not history.

For the method in more depth, see time series forecasting for demand planning. The principle is the same as in our post on machine learning for retail chains, applied to trade ordering rather than tills.

How much does it cost?

• Customer lapse model with a weekly rep list or CRM task: five to eight weeks
• SKU forecasting and suggested reorder quantities for one branch or category: eight to twelve weeks
• Rollout across branches with ERP integration: a further two to four months
• Quote conversion model: six to ten weeks, provided quotes and outcomes are recorded

On return, look at two numbers you already know: the value of stock you wrote down or cleared last year, and the gross margin from accounts that stopped trading with you. If either is small, a model will struggle to justify itself. For most distributors of any size, at least one is uncomfortably large.

When a model will not fix it

• The ERP cannot take suggested order quantities, and buyers will not rekey them
• Quotes are not recorded as won or lost, so there is nothing to learn from
• Stock accuracy is poor because goods-in and picking are not scanned
• Customer accounts are duplicated across branches, hiding real behaviour
• The real reason for lost customers is delivery reliability, which a list of at-risk accounts cannot solve

Several of these are operational software problems. Our post on software for wholesale and distribution covers them, and they tend to be the right first project when data is unreliable.

A distributor's machine learning project succeeds or fails on a very dull question: can the buyer accept the suggestion in one click?

Explaining the output to buyers and reps

Buyers have usually been burned by an ERP's auto-replenishment at some point and do not trust anything automatic. Show them why: recent demand, seasonality, the order frequency, the service level target. Let them override and record why. Reps, similarly, want to know why an account is flagged, such as "orders every 12 days on average, last order 31 days ago, basket down 40% over three orders". Those plain explanations get calls made.

Where we would start

At SpiderHunts, we generally suggest customer lapse detection first for distributors. It needs only order history, it produces a list a sales manager can use the following Monday, and the results show up in retained revenue within a quarter. Forecasting follows once the ERP integration path is clear. Both sit within our machine learning services.

Frequently asked questions