Inventory management and forecasting demand have always been a struggle for any company selling hard goods. Not only have bloated inventories historically been negatively correlated to profits, but also bullwhip effects have been cited as a main cause for economic downturns. On the flipside, under orders can mean unfilled orders and forgone revenue. In June and July 2015 we fell victim to bad inventory forecasting and were sold out of items causing a drop in our conversion rates to below 0.5%. Maybe more importantly we missed out on two months of crucial data that could be used for future inventory projections. At an early stage company, inventory issues can quickly exacerbate and there is low tolerance from founders and investors alike for inventory mistakes. With this in mind and given the general lack of historical data there is to rely on, late last year we began to use a triangulation of data techniques to forecast trends.
The formula relies on a three step process: decide which products to expand into and when, test the assumptions and figure out how deep to buy into those products, and then determine the effects to your current line.
1) Determine Where to Expand Using Publicly Available Data
The first piece of the puzzle is using publicly available data, mostly Google Adwords Keyword Planner (https://adwords.google.com/KeywordPlanner). We use this in two main ways: to judge overall demand and to match our order flow with demand seasonality. For instance, our original instinct was to produce a backpack for our product line, but Keyword Planner clearly showed that more pertinent demand existed for a duffel style bag.
Search data for gym backpacks showed 3,600 monthly searches, way below just about every combination of duffel bag searches. Searches for backpacks in general, while large, revealed intent to be way different from what our company was known for, so it was clear that a duffel was the correct direction. Furthermore, we used Makersights (discussed below) to test this hypothesis, which independently confirmed the Google results. During a product line extension, keyword planner becomes an essential tool to predict demand and to avoid overcommitting.
We also use Keyword Planner to determine our buy size and timing for a new product line. Without historical data to determine YoY trends, we can estimate demand based off of like products in our current line and adjust for differences in total selling period. The below screenshot shows “men’s swim” to have a much longer selling period than I initially would have thought, showing decent demand from January through August.
2) Determine How Deep to Go by Asking Asking Asking
The evolution of SaaS based systems has created a simplistic way to put customers through intricate A/B tests that will help forecast demand and price elasticity. We use a provider called Makersights, which allows us to test our product hypotheses against controls that we know historical demand for. This gives us insight into how deep to go into a new product line or if we should abandon a concept altogether. With limited capital and a limited ability to discount as a premium brand, these insights have proven invaluable in creating efficiency and avoiding that dreaded inventory bloat.
Makersights survey which allows us to build demand curves and discover purchase intent
With Makersights our controls have almost perfectly mirrored our historical data, giving us confidence in the platform’s predictive powers. We used the data to order the best performing colors of an existing product line. By selecting the colors with the highest demand, our models showed an 106% revenue gain over the next best performing assortment of color choices.
3) Understanding Cannibalization of Your Current Line
As a young company, one of the greatest challenges in forecasting centers around cannibalization. When launching a new product line, determining a cannibalization coefficient to attach on existing products can feel like the equivalent of throwing a baseball with your non-dominant hand - awkward and futile!
If our product line remained stagnant YoY, prediction would be relatively easy and we could get close with a simple growth rate multiple. Since our inception, however, our product line has grown from ~75 SKUs to 500, making forecasts much more intricate.
My early approach to forecasting cannibalization was to keep the category market share stagnant YoY (i.e. "shorts YoY would continue to make up roughly the same percentage of overall sales as the previous year) and assume that the introduction of a new SKU would simply rearrange the selling makeup within that category. However, after looking at the data pre-and post-launch of a new SKU, we found that the new SKU did infact not only cannibalize existing SKUs, but also surprisingly and consistently contributed to growing the category overall. The following graph shows sales of “short-sleeve shirts” consistently tracking with overall sales, but then abruptly diverging post the launch of a new SKU indicating growth in the category outpacing overall sales.
Similarly, when we launched our third short line the growth of our revenues from the “shorts” category outpaced our overall revenue growth by 23%. This trend has been observed and remains steady across most tested categories. With more options within a category, consumers shift purchases from outside a category to inside. This effectively splits cannibalization across multiple categories and minimizes its effect on the category with the expanded line. Now when we look at future product lines and YoY assumptions, we must factor in how growth within a certain category will not only affect the makeup of that category, but also the demand for its neighbors.
The difference between a successful product line extension and a bloated one is the intelligent use of demand data. Fortunately today, small businesses without a treasure trove of historical data are enabled to predict demand and ultimately better manage the inherent risk associated with a product line extension.