Expectations of AI in sales forecasting are often exaggerated

Many managers and administrators naively believe that artificial intelligence (AI) can finally provide accurate forecasts.
Instead of looking at historical data, it is better to rely on AI.
However, the reality is very different.

In reality, an AI system also learns from the past.
The AI may be able to recognize regularities beyond trends and seasonalities in the existing historical time series that traditional statistical forecasting methods do not recognize.
However, the ARIMA method is already very good in this respect.
Like any statistical approach, the ARIMA method does not reveal which specific factors influence the forecast.
However, an AI cannot do this either if it is only provided with historical time series from customer orders or outgoing stock data and no separate information on influencing factors that have had an impact on historical demand behavior.

If you have(prerequisite 1) assumptions about which external factors could influence market demand, you can go one step further and take these into account very efficiently and easily in AI-supported forecasts; especially if there are several factors that interact.
However, AI cannot find possible influencing factors on demand behavior independently.
At best, you can use large language models such as GPT (“ChatGPT”) to make assumptions about which influencing factors could have an effect if you have no assumptions yourself.

In order to quantitatively assess the effects of the assumed influencing factors on the market demand for a product or product group and to be able to make statements about future demand behavior, the AI also requires corresponding historical series of figures(prerequisite 2) on these factors with which the AI can be trained.

However, the fact that such historical data is available is not enough.
A sufficient amount of data must be available(requirement 3) in order to be able to train the AI.
How much data is required depends, as is often the case, on the individual case.
To put it pragmatically and unscientifically: 100 data sets are not enough, 1000 would be better, a few tens of thousands would be best.
In many cases, however, it is not possible to find that much data.

Furthermore, data on the influencing factors must also be available for the future(requirement 4a).
Only if you know how the influencing factors will change in the future can you determine the effects of this change on market demand.
Where this is not the case, there is still a chance(prerequisite 4b) that the current and historical data of the influencing factors already contain a statement about the future development of demand.
This sounds complicated, but it is quite simple: if, for example, the AI were to recognize from the training data that an increase in an economic factor, e.g. the IFO index or the purchasing index, leads to an increase in demand with a six-month delay, knowledge of the current value of these indices is sufficient to derive a statement about future demand in six months’ time.
In fact, some of our analyses from customer projects show that certain constellations of several economic factors together for certain product groups of certain customers make it possible to predict the development of demand for these product groups over the next three to six months.

As an example to illustrate this, let’s take the influence of competitor sales promotions on demand for a company’s own products.
In the case of standard products that are offered by different companies and that can replace each other, this is often a significant factor influencing the market demand for your own product.
Gathering information about competitors’ sales promotions can be very difficult, if not impossible.
At the very least, it would be necessary to know the dates of promotions and sales prices, and ideally also the sales volumes.