Warehouse Slotting Optimisation With Data

Travel time is the hidden cost in picking

In most manual warehouses, pickers spend more time walking than picking. The slotting was set when the building opened, new products went wherever there was a gap, and the fastest sellers now live at the far end of aisle twelve.

Nobody decides this on purpose. It accumulates. A warehouse with 6,000 SKUs and 2,000 order lines a day can easily add several kilometres of unnecessary walking per shift, and the pick rate report quietly blames the pickers.

What good slotting takes into account

• Pick velocity: how many times a product is picked, not how many units ship. A product picked 300 times in singles matters more than one picked twice in pallets.
• Cube and weight: heavy items low and near the start of the path, small items in shelving or carton flow rather than wasting a pallet location.
• Product affinity: items that are frequently ordered together placed close to each other.
• Ergonomics: fast movers in the golden zone between waist and shoulder height.
• Replenishment: fast movers need locations large enough that they are not topped up three times a shift.
• Constraints: hazardous goods, temperature zones, security cages and the occasional product that must never sit above another.

The data you already have is enough

Slotting is one of the few operations problems where the data is almost always good, because the warehouse management system records every pick.

The weak spot is usually product dimensions. Missing or wrong dimensions are common, and they cause a clever slotting plan to put a large box into a small bin. A few days of measuring the top few hundred products is well spent.

Where optimisation and machine learning each fit

Assigning products to locations is an optimisation problem. A sensible heuristic, ranking products by picks per unit of space and filling the best locations first, gets most of the benefit. A proper assignment solver that accounts for affinity and constraints gets more.

Machine learning helps in two narrower places.

1. Forecasting future velocity. Slotting on last quarter's picks is slotting for the past. In seasonal businesses, a model that forecasts next month's pick frequency per SKU avoids moving garden products to the front just as autumn arrives.
2. Affinity at scale. Association rules or embeddings trained on orders find clusters of products bought together that are not obvious from categories, such as a particular filter always shipping with a particular pump.

For a smaller operation, historical velocity with a seasonal adjustment is fine. For forecasting method choices, forecasting demand with your own data goes into more detail.

Re-slotting has a cost, so do it in waves

Every move is labour, and every move risks a mis-slot that sends a picker to an empty location. A plan that moves 4,000 products over a weekend will produce chaos on Monday.

• Estimate travel saved per move and rank moves by saving per unit of effort
• Do the top few hundred moves first; they typically deliver most of the gain
• Schedule moves during quiet periods and update the WMS in the same step
• Re-slot fast movers monthly or seasonally, and leave slow movers alone
• Measure pick rate and travel distance before and after each wave

A slotting plan that is 90% optimal and actually carried out beats a perfect plan that requires moving the whole warehouse.

When slotting optimisation is not the bottleneck

Sometimes pickers walk a lot because the pick process is wrong, not the slotting. Picking one order at a time across a large building will stay slow wherever products live. Batch picking, zone picking or pick-to-tote changes often save more than re-slotting, and should be looked at first.

It is also a weak lever for fully pallet-in, pallet-out operations, or for goods-to-person automation where the system manages storage itself. If your constraint is packing or loading, fix that. Our post on finding the bottleneck in your operation is a reasonable place to start before spending on slotting.

How SpiderHunts would approach it

At SpiderHunts we would build a travel model of the warehouse from location data, replay six months of real orders through it, and calculate how far pickers walk today. Then we replay the same orders through a proposed slotting and show the difference, move by move.

That simulation is the proof. If the saving is small, we say so and stop. If it is worth doing, the output is a ranked move list the warehouse team works through in waves, and a monthly job that suggests the next set of moves. This is a mix of optimisation and data science, integrated with the WMS through exports or an API.

Frequently asked questions