Why is my sales forecasting not being run by Artificial Intelligence?

Forecasting is such a fundamental part of so many businesses, yet some of the methods used to generate forecasts range from crude to comical!

• What will we sell this year in category X? Last year plus 5%, obviously
• What uplift will the summer promotion bring in? We expect a 3x increase in sales
• But what did the last summer promotion give us? About 1.7x uplift…

You get my point. A lot of what is labelled as forecasting is little more than finger in the air guesswork; on which meaningful business decisions need to be forged.

Knowing the importance of forecasting, and how challenging it may be for many individuals, why is there not some all-powerful AI tool that can produce these numbers for us? Part of the answer is in scale and complexity.

Any successful AI application will have been trained on an unfathomable volume of data. Chat GPT 4, for example, is estimated to be trained on a dataset of approximately 13 trillion tokens (think “words”, except they’re usually fragments of words or short strings of characters within text). By comparison, if you had daily sales transactions for 5 years across 1,500 locations for a given item you would still only have 2.7 million datapoints to work with. To be on a level playing field, you would need a mere 24 million years of daily sales data.

Complexity of AI predictive models

Generally speaking predictive models have parameters or variables that can be adjusted to subtly alter the outcome of a given calculation. Sticking with the Chat GPT theme, it works with circa 1.8 trillion parameters to generate its predictions. The most common forecasting suites will give you no more than 20 dials to tweak; most of which will have very little effect on any forecast generated, but will make the software easier to sell!

Also, even if you did have more variables to play with, the trick is being able to set them according to the patterns within each specific time-series. Something that is difficult to do successfully in any great scale if you’re relying on human intervention.

Knowing the size of the prize in getting sales forecasting right

This disparity on approaches seems obscene. Surely there is a way to harness more data and crank more handles to spit out something masterful. Well, maybe not. The other major hurdle for AI forecast generation is the size of the prize.

Consider the following:

Your best-selling product generates between 1,000 and 2,000 units of sales per week. It has a reasonable shelf-life and is supplied to you by a very reliable supplier that delivers within 2 business days, 5 days a week, in minimum quantities of 4,000 units. How accurate does your forecast need to be?

You could be 100% out in your weekly predictions and still not be in trouble. So why put enormous effort into trying to improve forecast accuracy if you’re not going to feel the benefit? It can quickly become a law of diminishing returns.

Anyone for snake oil?

That is why, I believe, a genuine AI sales forecasting solution is hard to derive. And probably explains why some companies (not wanting to name names, but please search for “AI forecasting” and you’ll soon find them) claim that standard statistical models are AI, yet berate the statistical and mathematical approach as it falls down in the face of complexity. I can see no reason to take this approach other than to leverage the buzz of AI to flog a service.

Forecasting and inventory: how can Hatmill help you?

The first step, before any forecasting can begin, is to assess the “forecastability” of your business. We typically do this by categorising all items into a volume and volatility matrix, which allows us to apply forecast and replenishment strategies at a more meaningful level.

For example, items that fall within the “low volume, high volatility” category are usually unsuitable for forecasting due to their sporadic demand patterns. In such scenarios we would recommend a min / max replenishment approach, whereby orders are placed as and when stock reaches a minimum threshold, to replenish the stock back up to a pre-determined maximum.

Items with sufficient sales volume and predictability can then be further assessed to understand which forecasting approach is most suitable. Some factors to consider during this process include seasonality, promotional activity, national events, and sensitivity to weather; as well as whether the item should be forecast at a granular, or aggregated level. Depending on the characteristics of an item’s sales there are different algorithms and methods better suited the handling these variations

This article is too brief to explore these topics in detail, but if any of this rings true for you and your business then please contact us to find out more.

Forecasting and Inventory