What history for a reliable sales forecast?

by Rupert Schiessl

#Prévision des ventes #Historique de donnée

How does a data history ensure the reliability of a sales forecast? How much data is needed for a machine learning model?

  1. The Role of History in Sales Forecasting
  2. What is the ideal history for a reliable sales forecast?
  3. The limits of history in sales forecasting

In a previous article, we discovered the importance of quality sales forecasting and the most effective methods for this exercise. Indeed, sales forecasting enables better decisions and reduces contextual risks. While the advantages of the learning machine are not negligible, it must nevertheless, and above all, rely on relevant data to meet the initial objective.

The Role of History in Sales Forecasting

Sales forecasting allows companies to anticipate future demand in order to make diligent management decisions, such as optimal inventory levels, investment in manufacturing equipment or energy, personnel planning, etc.

The decisions we make in our daily lives are strongly based on our past experiences. Our observations have allowed our brain to establish decision models that guide us in our actions. The same is true for sales forecasting, which must be largely based on historical company data, such as receipts, prices, promotional approaches or point-of-sale traffic. And like the decisions we make in our daily lives, forecasts are derived from a combination of patterns learned in the past and parameters observed at the time of the forecast.

There are many techniques for forecasting sales. In addition to the many so-called "naïve" methods (e.g. the prediction is equal to the average sales of the last 30 days), there are regression models (e.g. linear), which consist of establishing a correlation between explanatory variables and a target variable in order to deduce a trend. While these correlations can yield interesting results in a number of cases, this method has its limitations when it comes to gradually narrowing the gap between the historical points and the forecast generated by a model. In other words, the errors of the past form the errors of tomorrow. It is in the face of this limitation that the learning machine and self-learning methods show their usefulness.

Machine learning models are trained through thousands (or even millions) of iterations. At the end of each iteration, the algorithm evaluates the gap between the forecast generated by the model and the actual data from the history and attaches a "penalty" to this gap. This penalty minimization technique has the great advantage of being able to take into account very precise business objectives, such as maximizing sales and margins or minimizing inventories or out-of-stocks. Only when the minimum penalty is reached will the model be considered optimal and be able to deliver the desired predictions. It is thanks to this operation, oriented by precise objectives, that the algorithm will be able to understand, for example, that certain recent data, or those coming from the same point of sale, are more important in the prediction than older data or those coming from more distant points of sale.

What is the ideal history for a reliable sales forecast?

By definition, a real-world data history is full of fluctuations (and we will only talk here about fluctuations related to the natural activity of the company and not those due to poor data quality recording). These fluctuations may, for example, be linked to a notion of seasonality (mainly annual, monthly, weekly or daily). This is the case for seasonal products such as sunscreen or soup, but also for products linked to a calendar event such as foie gras at Christmas or school supplies. Some data may be structurally fluctuating, such as promotions or sales. Finally, sales can also be impacted by multiple external factors (demographic or economic changes, a change in store concept, a pandemic, etc.).

The algorithm must therefore have sufficient history to identify these fluctuations in the available data and to differentiate their random and intrinsic components. In general, one can expect a satisfactory consideration of seasonality effects from a data history exceeding 2 years. This temporality allows the integration of growth, seasonality or promotion effects in prediction models.

The limits of history in sales forecasting

An important limitation of forecasting is the fact that models can only predict what they know. It is indeed possible that some external factors may not be taken into account in the modeling, whether this is due to reasons of data accessibility, which in some cases may be difficult to extract from internal systems or to collect, or simply to the non-existence of certain data. For example, a promotion in a point of sale located in a shopping mall will be strongly impacted by the density of traffic within the same point of sale. However, if the store does not have a traffic measurement system and if the mall does not share its attendance figures with its landlords, it will be impossible to use this essential data in the forecast.

Also, as we discussed in a previous article, it is also possible that a paradigm shift may lead to the historical being no longer representative of the present. For example, the health crisis linked to covid-19 or prolonged strikes may abruptly change the consumption habits of customers. It is therefore necessary to provide the model with variables that allow it to identify this paradigm shift.

Finally, an important limitation to the use of history in forecasting is the absence of history for some products. This situation can occur when launching a new product, opening a point of sale, creating a new concept or changing sales conditions.

Take for example a promotional campaign offering a 50% discount on a bottle of milk. If the point of sale has never made this discount before, it will be difficult for the model to determine how many units will be sold during the promotional campaign and, without providing additional information, the algorithm will probably underestimate the quantity sold during this promotion.

To overcome this lack of history, complementary algorithms can be used to estimate proximity to other products with a richer history. This can be done simply by using the existing product nomenclature (families, sub-families, brands, etc.) as well as product properties (use, price, weight, color, etc.), or in a more sophisticated way, by matching products with similar customer behavior.

Want to know more about using the learning machine for sales forecasting? Ask for a demo of our platform now.

Be informed of the latest news

Receive our latest news directly in your mailbox