Bias in Amazon Forecast
Updated: May 2, 2019
How biased is Amazon’s forecast?
As most people know, Amazon provides first party sales a weekly forecast of expected ordered and shipped units. Most Vendor Central users also know that Amazon’s forecast trends more than a little to the optimistic side. How bad is this bias? Does it actually effect you in any way? We took a quick look at the level of bias in the Amazon forecast and how that effects Amazon’s ordering behavior.
We saw a 15% inflation at 1 week out
In one case, for a specific customer in a specific industry (that I am not disclosing), we observed a greater than 15% error a week out when looking over the average error in the last 4 weeks. This was the average error for more than 100 products. The total bias was always that Amazon was going to sell more than they did. For example, if Amazon forecasts they will sell 5000 of your products next week, more than likely they will sell less than 4250. This specific case may not represent all of Amazon’s forecast but should give vendors and sellers pause. Does this match other peoples’ experiences with their Amazon forecast?
How do we think Amazon’s forecast works
Our understanding of Amazon’s forecast is that they use glance views as a forward-looking driver of interest in a given ASIN. They then convert product interest into sales by a conversion rate and have some discount for expected buy box win percentage. Our speculation is that there is a category-based model with seasonality that predicts glance views into the future. The exact model used is probably very different by category.
Other forecasting issues with Amazon
The most frequent complaints we have heard regarding Amazon’s forecast are that it doesn’t change when prices change and that it doesn’t properly account for products going Out of Stock (OOS). The changing price correction is a really hard problem to solve as you need to understand the price elasticity of the ASIN. The problem if you don’t account for it at all is when ASINs fall in price by 50% your previous forecast assumptions go out the window and you end up in an OOS situation. Similarly, if the price goes up by 50% the product doesn’t move, and Amazon doesn’t reorder from you. Not accounting for OOS in ASIN history is where Amazon’s forecasting predicts a dip in future sales based off a past time period where you had an extended OOS ASIN. I believe this is a more historic problem than a current one with Amazon, but that could be a correction in the category models I have been following as opposed to a general fix.
What this means for Vendors
Ultimately, the sell-out forecast drives the orders Amazon makes with vendors. If they can’t forecast (or communicate the forecast) properly this causes bullwhip effect problems in the supply chain. Biasing a forecast causes the vendors pain in that Amazon frequently over orders and then waits too long to order again, resulting in a cycle of in stock and out of stock products. Not accounting for price and OOS problems further exacerbates the issue by not reacting ideally to changes. There is not much vendors can do about Amazon over ordering, however they can keep their own sell-out forecasts and better predict when and how much resupply Amazon will need to minimize OOS which leads to more sales and profit.