Is AI’s impact on demand forecasting more hype than reality?
Through a special arrangement, presented here for discussion is a summary of a current article from the blog of Nikki Baird, VP of retail innovation at Aptos. The article first appeared on Forbes.com.
The forecast error in retail is as high as 32 percent, according to some estimates. Will artificial intelligence (AI) technology do any better?
AI promises to change the way demand forecasting works in retail in six key ways, but those promises include a bit of hype:
- AI can examine a nearly unlimited number of causal factors simultaneously. AI’s accuracy requires the right kind of data.
- AI can apply analysis to every granular SKU, across every granular location where it will be sold. Problems arise when you start getting into very sparse or highly intermittent demand history.
- AI can apply advanced algorithms, like neural nets, to create new methods of forecasting. A wide enough variety of use-cases and data sets have not been tested to say that a neural net approach to forecasting should outright replace traditional models or even AI models that take a different approach.
- AI can select the right forecasting model to use for each specific circumstance. Where things get interesting is when you start looking at the next place where AI can be applied to demand forecasting, in the attributes that are used to identify when different products should use similar models.
- AI can identify when specific causal factors deteriorate in their contribution to the forecast and replace them with new, more important causal factors. Some AI-assigned attributes are not easily expressed in language that humans can understand. If misunderstood, AI’s recommendations can drive a company in the wrong direction.
- AI can react to changes in demand assumptions much more quickly than a human can. My sense is that it’s still an open question as to whether AI is really good at this or even if it’s needed.
Retail forecast error is high and anything that reduces it not only drives value on its own, but increases value in other levers too, like inventory levels, inventory turns and margin. However, it’s helpful to back away from using AI to mean some kind of nebulous future and dive into how, specifically, we might get to that future. It’s even more valuable to understand which pieces are within reach — and which are not.
Discussion Questions
DISCUSSION QUESTIONS: Do you see AI significantly transforming demand forecasting for retailers? What limits may the technology face when applied to forecasting, and what don’t we know yet?
Poll
BrainTrust
Paula Rosenblum
Co-founder, RSR Research
Gib Bassett
Director, Solutions Marketing with Alteryx
I recently had dinner with a group of CIOs who said their demand forecasting was all over the map. When I asked about AI, they said their merchants “don’t trust the machines.” As long as retailers are willing to accept mediocre results, we’ll live with the consequences.
Not sure how much AI is used in forecasting yet. The potential is there but just like everything AI, it’s early days.
There is a lot of promise but it hinges on a couple of key points:
AI looks at data without any emotional or sentimental influence. If the numbers are accurate, the data should be, which means the calculations AI makes should be accurate, as well. The key is to give the computer the right information. I’m seeing results that are improving logistics, distribution, inventory and more.
We need to be careful not to confuse AI with automation. Automation of things that can be automated in forecasting is wonderful. Letting a machine choose which attributes to emphasize at any given moment often leads to model over-specification – you get a great answer today when you build the model, not so much in the future (sometimes the near future).
AI relies on backward-looking data. For basic and commodity products that may serve for forward-looking models. In apparel, forecasting means best-guessing on changing and evolving variables. Extensive testing helps but testing does not create pitch-perfect models. Some forecasts will involve a high degree of “knowns.” Many forecasts will involve managing both “knowns” and “unknowns” — guessing. Human skill and judgement make the difference.
As time goes by, more and more applications of all sorts will be infused with AI and ML. Demand forecasting is just one piece of the puzzle.
The key is to determine how much personal data is permissible to use in AI-infused marketing and forecasting applications. That has yet to be sorted out.
What’s the old saying? “Garbage in, garbage out.” AI’s accuracy requires the right kind of data and the right kind of questions. Given the current margin of error on retail forecasting it’s fairly safe to say this is not being handled well by retail today in all cases. For AI to emerge as the driving force in retail demand forecasting it will be important to ensure data from all parties (retailers, manufacturers and other vendors) can be coalesced into one massive data set — this in and of itself may be too much to handle.
AI, like any system, only generates the quality of insights that it is fed. If you don’t trust your data then yes, there will be questions in the results. It is true that the more good data you feed AI, the more accurate the forecasting will be. We have seen a huge retailer generate double-digit percent improvement in demand forecasting via an AI tool. That performance was unprecedented.
Bottom line, if you have two years of reliable data, you can see results that are very impressive with the right application of machine learning.
The key to AI/ML is, what is AI solving for? Algorithms, to be effective, must go beyond computer and data scientists. A subject matter expert must be part of the solution to guide engineers, etc. to an understanding of what to solve for. To be successful, AI algorithms must be agnostic, stripped of subjectivity.
AI or advanced analytics of any form for demand forecasting or any use case continues to be a confusing topic for retailers, consumer goods companies and really any industry. Viewed as a one and done point solution or one time effort, it’s certainly possible that the hype exceeds the reality. The challenge companies have is developing analytics as a competency, and having the ability to test, learn and improve. The reality is the analytics leaders are using AI to improve forecasting because they make the investment in the right people, processes and technologies. For any other company they need a plan beyond looking at a single use case with a fixed outcome. The reason this is so challenging for retail and CPG is because of silos, decades-old processes and challenges to meet earnings quarterly. It’s a hard problem, but IMO it’s one that retailers thriving in five to 10 years will have overcome.
My experience is that merchants NEVER trust the “black box.” The so-called “black box” is MATH. Almost every set of algorithms in use today are a subset of math that is over 100 years old. While it is true that lumpy demand, new product introductions, seasonality, etc. are all challenges, new math can’t hurt!
Another tool, another arrow in the quiver will help innovative retailers survive the coming attractions. Every time I have been engaged in my consulting career to solve an inventory problem, the root cause has been human intervention, changing the forecast, using gut feel. As retail architectures mature, allowing us to have a real-time retail environment, AI will be an enabling technology that separates one from the pack.
In addition to forecasting, AI is dramatically improving fulfillment decisions. Innovative new technologies can even predict where customers will be in the next hour or next day based on historical patterns. AI is also helping retailers make better decisions on which store should fulfill an online order. While traditional logic would select the product from the closest store to the consumer, with machine learning techniques, retailers can assess the value of inventory in each store to make smarter fulfillment decisions. For example, if they can identify/predict that the item in inventory at the closest store will likely sell at full price, but the same item at a different store location will likely result in overstock and markdowns, the retailer can ship from the further store and maximize total profits.
Like almost everything in life, I think it important to note that one size does not fit all. Different categories require in many cases completely different approaches, even with AI. Long lifecycle products can combine historical sales peaks and valleys with causal factors that can be predicted with a high degree of accuracy. Short lifecycle products need a deeper look at outside information like weather, fashion trends and social buzz. This is especially true of using AI/Natural Language Understanding to look at outside factors that we have not looked at in the past. AI in all its different form factors is proven to improve forecasting by a factor of 10 over what we were able to achieve in the previous 10 years before AI. The longer we wait the further behind we will get from those that are doing it.
No one has a consumer product forecast that works. This is true today as it was years ago. All the research shows the same thing. External factors have significant impact on sales. In our many studies, weather and competitive active have significant impact some times. This variance is classified as the random effect. The real question is why anyone is spending time for forecasting today. We have cheap real-time computing to monitor the entire supply chain. Stop demand forecasting just use demand to run the operations.
AI will significantly transform most aspects of our lives. By 2045 (or so) a $1,000 USD processor will buy you the processing power equivalent to all human beings alive — about 9 billion people. It’s hard to fathom the kind of impact that’s going to have on our businesses or our lives; however, I imagine shoring up at 32% forecasting error will be positively impacted sooner than later.
Dues this sound familiar? Someone invents a new technology — maybe it’s the wheel, the steam engine, AI — makes modest claims about it. Someone else comes along, makes all kinds of outlandish claims about it, attracts intention from the (generally uninformed) media, and all sense of proportion is lost.
Will AI improve forecasting? Certainly, particularly if it’s used by people who know what they’re doing. But will it “transform” forecasting? No: we won’t be able to know tomorrow today (and I’m not sure we should want to).