Machine Learning and statistical models are complementary

Sales forecasting methods based on artificial intelligence, including Machine Learning and Deep Learning, are now nuggets that companies can add to their arsenal of methods and techniques to develop their sales forecasts.

How can Machine Learning predict sales ?

With a large amount of sales data over long periods of time, it is possible to identify patterns of buying and consuming behaviour of individuals.  Machine Learning algorithms take into account, on the one hand, various external and internal factors that have an impact on sales. These factors include promotions, customer loyalty, positioning compared to competitors, etc. On the other hand, they take into account the likelihood that the sale process will or will not be carried out.

Machine Learning vs Statistical Methods ?

Some articles claim, without any nuance, that machine learning can produce more reliable forecasts than those obtained by statistical models. This remains to be demonstrated by comparative studies (or “benchmarking”) such as some papers published in the International Journal of Forecasting (IJF). Authors have tested statistical techniques and methods on several time series, including Machine Learning and Deep Learning. One of the conclusions I can draw at this point is that in some cases, Machine Learning provides more accurate forecasts. In other cases, statistical models are more efficient at predicting. 

When statistical data is not well structured and exhibits significant erratic variations, Machine Learning can provide more accurate forecasts than statistical techniques. On the other hand, if the series are well structured with some stability, statistical models provide more reliable forecasts. Another advantage of these is that they can handle accidental phenomena and significant variations with conditions to model them correctly. For example, if a data set is disrupted from time to time by accidental events (strikes, high-impact climate events, etc.), it is possible to take them into account in a Box-Jenkins model with intervention function. Ditto if this statistical series is impacted by significant variations in the market structure for example (competition, regulation, …). The Box-Jenkins model with a transfer function, for example, allows such changes to be included in the modeling so that the forecast takes into account new events and effects.

To conslude, it seems to us that for a company that has a wide variety of data in terms of structure, the use of the two approaches mentioned above can be complementary and beneficial.

PREDICONSULT offers two courses in Machine Learning

PREDICONSULT is offering two one-day courses each. The first, entitled “Predicting with Machine Learning,” aims to present Machine Learning models for time series forecasting.  Indeed, the databases available to forecasters are increasingly large and sampled at a high frequency. Knowing how to process this type of data requires appropriate methods such as Machine Learning.  It will detail the models with many applications on real data.

The second course, entitled “Machine Learning: Theory, Approach and Case Study” presents a broader view of Machine Learning. It is very practical and is aimed at all those who wish to be operational quickly in Machine Learning.  At the end of this training, participants will be able to conduct an in-depth analysis of their data from start to finish, use the statistics to make the data speak, know how to do the most common statistical tests and be able to interpret them.

Submit a comment