Chapter 1

Demand Versus Sales

To grow sales a good place to start is with demand, because:

Demand  =  Sales + Lost Sales 

Every retailer has lost sales mostly caused by out of stocks in stores or the online warehouse. In one sense, these are the easiest sales to recover, because you don’t have to win them from competition. You just have to minimize out of stocks and you do this by factoring the lost sales into your sales planning and forecasting.  

The second thing to appreciate is that demand is demand at a price. As a sweeping generalisation,
if you reduce the price of something, you mostly sell more.

So let us examine these two aspects of demand in more detail.

Demand Rather than Sales

Lost sales in retail ranges from about 6% of sales in grocery to more than 20% in fashion apparel. This range is so wide because grocery forecasting errors are smaller, as almost all products are continuity items and have a lot of sales history. This allows advanced forecasting models to achieve greater forecast accuracy. In fashion, a good line may sell one SKU a week in a store and you often only have this season’s history, so you get higher forecast errors.

With lost sales in this kind of range, if you could reduce your lost sales by half, this would be a 10% sales increase in fashion and a 3% sales increase in grocery. Hence it is worth making a serious effort.

The first step in the journey is to capture daily out of stocks by store by SKU. In practice, when you estimate demand using sales plus lost sales, you may reduce the theoretical lost sales number by a factor (often 50%), as some customers experiencing an out of stock will buy a substitute instead. In your database, keep your sales numbers and lost sales numbers separately, as planning or forecasting things like returns can only be based on actual sales. The calculation of lost sales is done in units and then multiplied by the prevailing retail price at the time of sale to get a money number. The price also needs to be stored by SKU by time as it will most likely change as things like promotions take place. 

Apart from factoring in lost sales, what else can you do to improve forecast accuracy?

There are a number of approaches you can use in combination: 

  • Forecast at the right level in the merchandise hierarchy. Forecasting at store SKU has very high error rates.  Some surveys have found an average error of 50%. Much better is to forecast at warehouse level using the aggregate of demand from all the stores served by that warehouse. In order to replenish individual stores, you will then need a good allocation process. 

  • Most stores carry a certain number of weeks sales as their inventory level. A fashion retailer might have 8 weeks sales in store, for example. If you forecast a rolling 8 weeks, at store SKU, the forecast will be more accurate as some of the weekly plus and minus errors will cancel out.
  • There are times when the recent history is of little use. An example is during the Covid store lockdowns.  In these circumstances, it is best to go back to an earlier year where the history is more “normal” to use as the baseline for demand forecasting.  In this case, you apply the lost sales for the year in question to get estimated demand. 

  • Modern machine learning based forecasting systems also do a better, more accurate job. These systems use advanced statistical techniques like regression analysis. They include a variety of other factors in addition to demand history to adjust for differences, such as promotional calendars, weather, holidays and so on.

Demand at a Price

Most products have a price elasticity. This means that as the price falls, sales increase. This image shows an example of a price elasticity graph.

All products have an initial retail price which is the price it first goes on sale at when it arrives in the stores.  The gross margin this price gives the business is known as the intake margin or initial margin. Periodically, the product may be reduced in price as part of a promotion. If so, it will return to its regular price when the promotion ends.  If it is a seasonal product, say a spring summer product, and it does not achieve its planned sell through, it will be marked down at some point in the season. Precisely when this happens will depend on how badly it is missing its sell through plan.

Clearance markdowns like these go down and may go down again. They do not ever go up. As prices are reduced to make sales increase, an eye must be kept on the achieved gross margin to make sure that the benefits of the sales increase in units are not wiped out by the reduction in gross margin. In branded goods, the brand owner may contribute to the markdown cost, but this does not happen with private label or own brand goods.

If a product is in short supply, you can use the price elasticity curve to reduce demand to more nearly match the supply available.  Grocers do this naturally with fruit and vegetable sales. The higher retail price will help protect money sales even though you can’t sell the same number of units. In non-food, the best time to adopt this approach is before the season starts, if you know that supply will be limited.  Customers will be more resistant if you change the price upwards in season.

Products With No History

A lot of seasonal products have no previous history of relevance as they are new items or new flavours, styles, etc.

There are two aspects of dealing with this situation:

  • Getting the best seasonal profile  
  • Deriving an initial seasonal estimate of sales to apply to the profile curve. 

This chart shows a sales profile and the sales in each time period are expressed as a percentage of the sales for the whole season (which may include sales into the first few weeks of the next season to clear carry over stock). The shape of the curve is the important thing and you may adjust the shape of the curve to reflect a different promotional calendar or other important change. As before, this profile should be based on estimated demand not actual sales. An example is Easter, which moves the week in which it occurs every year. If you know that Easter will be 2 weeks later next year, you can move the Easter week percentage later by two weeks and make last year’s Easter week a more normal weekly percentage for that time of year.

There are essentially two ways to estimate the seasonal total. 

Getting the best seasonal profile

The first is to build the profile based on the total sales for all products in the category, sub-category or class. The level in the hierarchy should be selected as far as possible to be representative of the product and not include other dissimilar products. Expressing the curve as percents of the season as described above is step one.

Then rank the individual products highest to lowest by total seasonal sales and pick product that will most closely resemble the new product expected performance based on the buyer’s judgment and market knowledge. The sales for this product will be applied to the profile curve to create a week by week demand forecast. 

This process can be improved if you have a rigorous approach to assigning attributes to products. Ideally, you should assign product attributes to each new product as you create them in your systems. Attributes can include things like: 

  • Price
  • Silhouette or product shape
  • Skirt length
  • Cuffs (turnups) or no cuffs on trousers
  • Double or single cuffs on shirts
  • Wear with cufflinks
  • Fabric composition
  • Fabric treatments like wrinkle free
  • Colour
  • Fit
  • Flavour
  • Type of wood in cabinet furniture
  • Etc.

The list does not need to be lengthy for each individual product but should include those likely to be most important.  

You can then use a balanced scorecard approach to calculate a weighted average attribute score for each product and select the product with the score closest to that for the new product. The higher the score, the better the historic product serves as a guide.

Having set up attributes, you can run some trial calculations on your history to find out which attributes are most significant in any one product category.

Deriving an initial seasonal estimate of sales to apply to the profile curve

The other approach if you have rigorous attribute definitions is to calculate the balanced scorecard as above, take the seasonal total for the most representative product and use the demand history for that specific product to create a sales profile. This could introduce bigger errors though, as developing the profile shape at higher levels in the hierarchy smooths out some of the random noise in the data.

Using these process variants, in suitable combinations and supported by appropriate software will greatly reduce forecast errors and this in turn will result in higher sales in both units and money.