AI in Demand Forecasting and Planning: From “Is it really AI” to “What makes it AI”

What defines AI in Demand Forecasting and Planning. Is it the process, technology, or outcome? Especially when there are many offerings that talk about using AI/ML in this space. Customers often ask the question “Is it really AI”. So how should one think about AI in this space?

Appreciating the complexity and nuance in the Demand Planning and Forecasting process

This question, “Is it really AI?”, comes up a lot. Demand Planners have been using “statistical” forecasts for many decades. So, what separates an AI based forecast from just another better “statistical forecast”?

It is a valid question especially in an environment where there are a lot of claims about AI/ML that fall short.

In common applications AI is usually associated with “human like” behavior such as speech recognition, image and video recognition, natural language understanding and generation, autonomous navigation, etc. Common examples are Chatbots, Alexas, Robots, Self-driving features of cars, etc.

So, what defines “AI” in Demand Forecasting and Planning?

Especially since Demand Forecasting and Planning is a complex and nuanced activity.

It is not just about producing statistical forecasts.

If one looks at the process, there are a series of interconnected steps:

1. The process of demand planning starts with collecting, reviewing and cleaning historical data. The goal is to maximize the quality of information that is being used as well as clean the history of known one-time anomalies /outliers and bad data that are not systematic and will not re-occur in the future. While some of the data cleaning process is required, a lot of the data-cleaning is done to satisfy the requirements and limitations of the downstream statistical algorithms.
2. Once the data cleaning is done a base-line statistical forecast mostly based on traditional time series methods. Significant effort is put into to generate these forecasts — using different models to generate the forecast and picking the best model to suit the requirements. Depending on the company, these tasks are either done by the Demand Planner or an in-house or outsourced statistical forecasting team.
3. Baseline forecasts are just a starting point. The real work of enriching these forecasts is undertaken to get the forecast to improve as much as possible. Manual enrichment is the most value-additive step of the forecasting process. This is where the real Forecast Value Add of the demand planner and planning process is seen and is also one of the most time consuming and error-prone activities.

This takes the form of:

• Taking inputs from sales & marketing teams
• Considerations such as promotion alignment with SKU-DC level, managing turns, customer inventory pipe-fill, competition impact, etc.
• Accounting for distribution changes (gains or losses), DC and customer realignments, phase-ins and phase-outs, new products, trial orders, promotion timing changes, etc.
• Assessing forecast roll-overs, DC splits, MAPE and Bias trends and exceptions, etc.
• Agreeing on consensus numbers at a category/brand/national level based on targets
• Top down and bottom up reconciliation of the numbers from the highest to the lowest item-location-week/month.
• The understanding of key external factors such as weather, category trends, macro-economic trends, etc.

Exceptions are one of the toughest areas since these represent situations that born out of changes from the routine process. However, many exceptions are also repetitive in nature, but are called exceptions since the statistical forecasts are not able to account for them.

The output of the above process is the final “operational forecast” or “enriched forecast” that drives the upstream and downstream execution.

Needless to say, the above process takes a lot of time and repeats itself depending on the frequency of planning and re-planning in different organizations. During the cycle Demand Planners do a number of other tasks like participating in meetings such as Consensus, Service-Level, S&OP, etc.

It is important to note that while there are individual demand planners that generate, enrich and enter the final enriched /operational demand forecasts it is really an output of an organizational planning process that the demand planner spearheads towards generating and stamping the final forecasts.

How to think about AI in Demand Planning and Forecasting — Alan Turing to the rescue!

So, when we ask the question, “Is it really AI” for “AI in Demand Forecasting and Planning” — which part are we referring to? The technology, the process, or the output? What defines “AI” in Demand Forecasting?

To address this question, we decided to go back to how AI is defined.

There are many ways of thinking about AI.

First, there is the distinction between a Narrow/Specific AI applied to a use case vs. General Artificial Intelligence, that can be applied across many use cases.

There is no confusion there since we are talking about the specific use case of Demand Forecasting.

Second, there is the aspect of the kind of algorithms and factors that are getting used. Most AI applications use a mix of advanced algorithms that come from the Deep Learning, Machine Learning, Computer Vision, Natural Language Processing/Generation, etc. However, is the use of such algorithms sufficient to define “AI in Demand Forecasting”?

Next, we turned to a really attractive definition by Alan Turing and his well-known Turing Test for inspiration. The Turing Test was articulated by Alan Turing in 1950. It is a method of inquiry of a machine and tests a machine’s ability to exhibit intelligent behavior indistinguishable from a human being, and thus the name “Imitation Game”.

Let’s say I was interacting with a computer/machine at the other end without knowledge that I was interacting with a “machine”. And in my interaction, I could not sense that I was interacting with a machine then that machine would qualify as a “machine that can think” or as we choose to say qualify as an “AI”.

Although many believe that the Turing Test in its original framing might have lost its relevance in today’s world, we think the simple intent behind the Turing test is still relevant (although it does have its limitations).

So, is there a “Turing” like test that we could devise for the specific use case of Demand Forecasting? And what behavior would such a machine need to exhibit to qualify as AI and not just another statistical/machine learning algorithm?

Note, we are using the Turing Test only as inspiration in our endeavor to have criteria to answer the question “Is it really AI” in the context of Demand Forecasting and Planning.

The beauty in the concept of the Turing Test is that it does not put specifics on the subject of the interaction. It also does not put any conditions on how the machine was built or trained (supervised, unsupervised, etc.). It merely focuses on the output of the machine and whether the output of the machine and how it seems to get to the output (the thinking process) is indistinguishable from the human being when posed the same questions.

What is also interesting is that it does not put any onus on the performance measurement / usefulness of the output. The output is simply to be compared only against a human generated output. Now, since then there has been significant research on the area of human cognitive bias and that humans might not make the best judgements and decisions under various circumstances. Especially in the world of demand forecasting and planning where human bias in planning is an active area of research and study.

We can take note of this and can add a condition that the output should perform as well as, if not better than what the existing demand planning process would produce for the use case at hand.

E.g., in the case of Demand Forecasting the performance criteria is usually a comparison on the machine generated forecasts vs. demand planner/planning process generated forecasts in a “blind test” (prior to the actuals being observed) against the actuals that were observed.

One important point is that to ensure a reasonable comparison in the “blind test” the “AI” needs to have access to the same information that demand planners/planning process usually have. In many cases maintaining an apples-to-apples comparison between the demand planner-process generated forecasts and machine generated forecasts becomes difficult. Especially since demand planners might have access to last minute information, inputs from sales, marketing, business targets, competitive actions, etc. that the machine might not have to access to since that data might not have been digitally captured in any system.

With the aforementioned in mind, and assuming that both the demand planners/planning process and the machine have access to the same level of information, one might come up with a candidate list of criteria that an “AI” should exhibit in terms of the “output”:

1. The demand forecasts and plans generated by the “AI” would be consistently equal (if not better) to the demand forecasts generated by the existing demand planning process
2. The “AI” would be able to correlate and quantify the impact of different activities such as sales, marketing, promotions as well as external signals such as weather, macro-economic trends, competition, etc.
3. The “AI” would be able to apply intelligence and make decisions to handle exceptions for both high volume and long tail product-location combinations
4. The AI would be able to incorporate human inputs and judgement while removing any specific bias
5. The “AI” would be able to apply scenario-based and probabilistic thinking to make forecast adjustment decisions
6. The “AI” would be able to dynamically adjust forecasts based on latest demand patterns, internal and external signals

The above is only a candidate list. In-fact, one could come up with criteria for AI for each sub-step/group of steps in the demand planning and forecasting process, as well as increase the bar on the performance criteria. E.g. the demand forecasts and plans generated by AI should be at-least X% better than the current forecasts produced by the demand planning process at the operational levels to drive significant business impact.

The above kind of AI might not necessarily be the kind we usually expect- Alexas, Robots, Chat-bots, Image Recognition, Speech recognition, etc. But it will nevertheless be an “AI” that can produce demand forecasts and predictions better than a demand planning process output while exhibiting “human-like” reasoning and judgment.

To be clear we are not talking about robotic process automation with automation of rules-based logic. It is not just about automating the process, but about applying intelligence, reasoning and judgement that a demand planner/participant in the demand planning process would normally apply.

Also, this kind of “AI” is not necessarily one that is designed to replace demand planners or the planning process. As most “AI” s in today’s world this AI would help amplify human and process intelligence to make better decisions, identify blind spots, as well as help in performance and cognitive feedback.

From “Is it really AI” to “What makes it AI”?

So, one might ask, how is this exploration at defining “AI” in demand forecasting and planning useful?

The concept and promise of “no-touch/touch-less” planning and forecasting has gained steam over the past few years and “AI” and “Machine Learning” algorithms have been applied to Demand Forecasting in the CPG industry. Each company has their own flavor of AI with different emphasis. However, we find many demand planning teams unable to differentiate between offerings due to lack of a framework on how to think about “AI” in demand planning and forecasting and what they should expect out of an AI offering.

Further, irrespective of the flavor of AI some of the key questions w.r.t. to demand forecasting and the endeavor towards “no-touch” still remains — does the “AI”/” Machine” generate demand forecasts/plans that are as good as or even better that those currently generated by the traditional demand forecasting and planning process. And, in that endeavor does it alleviate some of the challenges and pain points that demand planners and demand planning teams face.

As many offerings claim to be AI it is worth asking the question — “What makes it AI?”.

Contributors: Rajat Srivastav, Director — Product Management, Rohit Kumar, Director — Product Solutions, Ankur Verma — Senior Data Scientist, Yadunath Gupta — Senior Data Scientist, Siddharth Shahi- Data Scientist