Part five of our blog series on improved forecasting using cognitive insights
In our previous blogs, we’ve seen how a forecasting and demand planning assessment can give actionable insights that enable the planning team to focus on moving towards smart-touch forecasting and increase the quality of their decisions. Furthermore, we have some first results on the effect of decisional guidance: It generally improves forecast accuracy, but a planner’s willingness to accept the guidance can vary.
Just like there are many different kinds of people, there are many different kinds of planners. We will highlight three examples:
- An optimistic planner adjusts too heavily, typically in an upward direction, often decreasing accuracy.
- A planner can furthermore show anchoring behavior. The adjusted value is in the right direction, but the adjusted value is too close to the statistical forecast (increasing accuracy, but not attaining the full potential).
- Finally, an overreacting planner adjusts in the right direction but overshoots the actual demand.
A theoretical data scientist might now simply say: “Easy! We identify which kind of planner we’re dealing with, and accept or reject their enrichments based on a fancy machine learning model, thereby maximizing the accuracy of the final forecast”. This is, however, not a productive approach: It is pitting the human and the machine against each other, instead of empowering planners to make the best possible forecast. Instead of ‘human versus machine’, we could be maximizing the potential of ‘human with machine’.
So, why do we care about planner types?
Imagine for a moment that you’re talking to two planners, let’s call them Anna and James. Anna is a typical optimistic planner: She has a positive outlook on life in general and that filters through to the demand planning enrichments he makes. James on the other hand is an anchoring planner, and a bit more cautious.
In the real world, talking with Anna and James would be a very different kind of conversation. Similarly, if we want to improve their impact, a planning system should give very different kinds of feedback. To Anna, when she adjusts a forecast upwards, the machine may suggest: “Are you sure? Dramatic upward adjustments like this have shown to be overly optimistic in the past. Please consider both the size and the timing of the uplift.” To James, when he adjusts a forecast upwards, the suggestion may be: “Are you sure? If you have good reasons to increase the forecast, the uplift will probably be more than you think.”
‘Smart-touch forecasting’ through automated enrichments
The examples above are a natural first step. Once a planner has made their adjustment and the system sees a potential for significant improvement, we give targeted feedback. This approach puts the planner in charge and leaves it up to them to accept or reject the machine’s suggestion. This is a small first step to take: When the recommendations make sense and prove their value, over time demand planners will come to trust the machine.
That’s when it’s time for the next step: automated enrichments. Taking as input the time series data, historical enrichments, and other internal and external drivers, we can use machine learning techniques to automatically recommend enrichments. A planner can then focus on validating those suggestions.
With modern advances in the explainability of machine learning models, techniques such as Shapley values (e.g., SHAP), and local surrogate models (e.g., LIME), we can create an understanding of the behavior and outcome of the machine learning models. For instance, the planner can see what inputs, trends, and drivers have caused the recommendation. This further enhances trust in the algorithms and enables the user to make an educated judgment call on the validity of the recommendation.
In the current day and age, with a ‘war on talent’ in the job market, our aim should be to empower and engage our planners. With a step-by-step approach from cognitive insights in a forecasting and demand planning assessment, through personalized feedback, and finally automated enrichments, we can truly enable smart-touch planning. By giving our people the right tools and the right information, they can have a tremendous impact.