In part I of the publication series Strategic Insights, the potential benefits of extended statistics and machine learning for strategy and planning work were outlined. As a more in-depth introduction to the topic, Part II discusses ways of increasing the quality and reliability of forecasts in planning.
Every year, we provide forecasts on market volume, sales and earnings for our applications, products, customers and regions as part of complex planning rounds. In our volatile businesses, we often have the feeling that a glass sphere could serve us as well as our attempts at forecasting. The process itself is perceived by many managers as costly with too little benefit for their own area of responsibility. Too many different assumptions of many involved actors and the high complexity of our planning structures are a challenge.
The meaningfulness and usefulness of the forecasting work should not, however, be called into question despite all the difficulties:
Good planning work creates a robust basis for the pre-steering of our businesses and creates added value for all the players involved – especially in times of high complexity and dynamism.
Advanced statistics and machine learning paired with the right technology enable us to make a practicable entry into two subject areas with high potential benefits for our business.
If numerous planning units provide planning figures for volume, sales and earnings over several years at least annually for all their regions, application fields and product groups, then the existing time series and data points can quickly reach thousands. Who can look at all this in detail?
We need to find a way to automatically sift through all the data available, taking time into account. Each forecasting unit (e. g. sales development of a product group in a region) is examined with regard to its relative importance for the business, its volatility and its forecasting error.
If the condensed results from this automated filtering process are transferred into a forecast quality portfolio, it becomes clear where the relevant candidates for the business can be found for further investigation:
This relevance detection considers three variables over time:
By using quite simple statistical methods, we are in a position to check the plausibility of developments in our businesses, to discover overarching patterns in a large number of planning units and to sustainably improve the quality of our forecasting work. And all this without having to spend days searching in haystacks.
During the planning process we naturally focus on the future, historical developments play a subordinate role. What would happen if we had a digital assistant at our disposal that knew and always kept available all past figures, hundreds of time series, correlations and correlations?
This is precisely where the core competence of machine learning models lies. To recognize statistical correlations from a vast amount of data and to derive forecasts from them. Such models learn from our past and recommend probable scenarios. The machine becomes the perfect complement to the manager with his or her experience and fine feeling for the business. The greatest weakness of the model is at the same time the greatest strength of the human being and vice versa.
So-called predictive modeling helps us at different points:
The importance of forecasting work is too high to allow it to be completely left out of hand and for the machine to decide for itself. Ideally, humans interact with the algorithm (human in the loop). This includes interpreting the simulation results, deciding which values to include in the planning and which impulses to return to the machine in order to rotate learning loops and optimize the algorithm.
Assessing market and business developments has nothing to do with analytical-deductive science. It has to do with finding a balance between today and tomorrow, the inner and outer world of an organization. This requires experience, intuition, creativity and perception. Everything that distinguishes the human leader and no machine can replace. These intelligent, digital new possibilities will soon find their way into today's organizations. We should only make sure that the tail does not wag the dog at the end.
With a little practice and used correctly, we will make our planning work more reliable, efficient and stress-free for ourselves. We no longer speak of abstract ideas. In a current project, we have succeeded in beating the accuracy of human forecasts by a factor of 2 to 3 with mechanically generated forecasts of sales.
In the third part of the the publication series Strategic Insights we go deeper into the methodology and systematics of machine learning technologies and predictive modeling.
