Glass Ceiling in Demand Forecast Accuracy – Who is to Blame?

I’d be a very rich man if I had a dime for every time I heard a senior supply chain executive say to me “I need the demand forecast accuracy to be xy%”. After a short smirk, I sympathize with them – it appears that they throw everything at the problem to attain a higher demand forecast accuracy except that they hit a glass ceiling that no one knows how to break.

I have personally consulted with a dozen Fortune companies in the areas of demand and inventory planning and have talked to dozens more on the subject. I wanted to present my learning in this post and perhaps address, at a high level, a novel and relatively new concept that is taking shape. I think it could change the way our supply chains work.

I have already talked about the dichotomy between data science and business analytics in another post. So, at the risk of repeating myself – Business forecasting is OH-SO-NOT time-series forecasting. Why? Simply put – Business forecasting is tied to scorecard metrics that time-series forecasting has no visibility to.

Imagine this.

    1. You are a widget manufacturing company, and your CEO wants 20% revenue growth YoY. It’s been committed to the street, so if you miss it, bad things happen.
    2. Sales braces itself and lays down its strategy to achieve 4.7% more every quarter. They have two levers to play with – Pricing (~Margin) and Volume. They lay their bets – drop prices on ½ the portfolio (Portfolio 1) to get a disproportionately high volume to attain a 10% (aggressive!) growth on that piece; brace for a slower growth of 2% (passive!) on the other half of the portfolio (Portfolio 2) where they see an inability to compete with Asian products.
    3. Then they go to Marketing. Marketing comes up with a tactical plan that is more month-on-month. They blend in the promotions and campaigns, and the plan starts to look good. But their plans are at a product portfolio level.
    4. Supply chain takes the portfolio level forecasts and makes a week-on-week SKU-level forecast. This is where the magic happens!
    5. Supply chain, sales, and marketing folks talk. Supply Chain confirms capacity, supplier readiness, and their stocking plans. All systems are go on Day 0!
      Does sound like the ideal world, doesn’t it? Alas, one custom part has a lead-time of 12 weeks! Double alas, that is a critical part in the portfolio that Sales doubled-down on (Portfolio 1). Triple alas, the supplier has yield issues – their making process must be requalified and is going to take a few weeks.
    6. The supply chain, sales, and marketing folks talk. Procurement is happy to chase down the parts needed but Plan B is put in action – the second half of the portfolio that sales was sandbagging on is now the last resort.
    7. Without the disproportionate volumes on the former portfolio, margins got majorly dented due to price drops. Inventory on that portfolio vanished quicker than expected.
    8. With slower growth expectations, the supply chain was unprepared for Portfolio 2’s demand, and in a competitive environment, that is detrimental.
    9. Repair measures came in too slow and in trickles.

Read more: 8 Must-have Analytics Capabilities for a B2B Firm

At the end of H1, fiscal targets don’t look so good – EPS expectations are missed.

I know I just had a lot of supply chain executives relive their worst nightmare, so I’ll quickly jump to why I did that. I’m going out on a limb here – No time-series forecasting model can and will ever be able to capture any of the above occurrences – as often as they happen – even though every one of them is captured in the scorecard metrics for each function.

So, here are five things you need to know about demand forecasting:

1. Reliance on history

All time-series models are fundamentally arithmetic! It bothers me when people talk about these statistical models as if they are a person with common sense – they are not (I know. AI, right?). They are ratios and functions formed out of past data. History itself is an artifact of past wrong-doings. And past data does not know that your CEO committed to a 20% YoY growth for the next fiscal!

2. Reliance on accuracy

All time-series models try and minimize two things – Percentage Error and Bias.

Bias is tricky – it is the balance between going over or under the real demand. You could have zero-bias and still be just 40% accurate over time – that would be an insane problem for the supply chain. Bias is really not as bad an evil as people make of it. In today’s world, you are so much better off with positive bias in the short term rather than pure inaccuracy.

But every arithmetic model tries and minimizes it, thereby risking going under the actual demand. If this happens often enough, backorders easily pile-up, laying even more stress on the supply chain. Also, accuracy is a realization of your own inability to match demand.

3. Supply chain needs an operable forecast

Supply chain does not willingly generate its own forecasts at an SKU level (By the way, if you are Supply Chain and don’t generate your own forecasts, you should call me). It’s forced to do so because no one else with more information can.

Sales does not have the time and Marketing can only do a good job at a product portfolio level because that is how they go to market. Concepts on CPFR were born out of this predicament, and God knows it is applicable to but 25% of a practicing organization’s portfolio by volume. The other majority gets arithmetic projections of the past that have no sense for any of the aforementioned.

4. Plan for the worst and hope for the best

If you know that ½ your portfolio has a core component that is sole-sourced, you’d have a Plan B and Plan C ready no matter what demand forecast accuracy you work with. It takes small supply chain failures to whip through demand predictability across the value chain.

5. It is always too late when you realize your forecast accuracy is low

The time-series model itself (implying the coefficients, diagnostics, parameters, hyper-parameters if you like) had failed a long time before you realized that the accuracy is low. You just had no visibility to the governing metrics.

Read more: 5 Brilliant Digital Transformation Case Studies You Need to Read

And finally, why and how does business forecasting get you out of this predicament?

  • Business forecasts use time-series models in an informed way. Business forecasts for an SKU or a product portfolio are generated by an entity (person/partner/excessively smart application) that used the time-series model and augmented all the information within the various functions and built out an operable forecast. Business forecasts know about what the CEO committed to the street. It knows about the sales strategy and the risk in it. It knows that a Plan B is required.
  • Business forecasts do not purely rely on accuracy or bias. Instead, they work on the allocation of what I like to call “Penalty Budgets”. Business forecasts know that the organization needs to park more working capital in tricky supply chain scenarios despite getting crucified by Finance for doing so — a necessary evil.
  • Business forecasts are regenerated far more frequently than anyone else creates a forecast. They always lead the conversation between Sales, Marketing, and Supply Chain rather than participating. They are the “I” in Integrated Business Planning. A 14-days-ahead daily forecast to manufacturing followed by a weekly forecast for 13 weeks is standard. Except business forecasts don’t just take a monthly forecast and divide it by 4 to send out that weekly forecast. They are more rational than that.
  • Business forecasts forcefully use Sales as their eyes and Marketing as their ears on the ground. As the cliché goes – “ANY information two weeks prior is better than ALL the information post fact.”
  • Business forecasts GOVERN the time-series models that go into the process of making predictions. Governance is not the same as measuring accuracy. To measure accuracy, you already need an actual demand realization. If a model is poorly representing the real-world demand, parameters like the R-Squared or the Maximum Likelihood/Akaike Information Criterion for slightly more complex models are leading indicators of failure.

Read more: How an Australian Specialist Services Company Optimized Procurement Costs

Of all the business functions, I love the supply chain function the most. Having spent so much time with them, I have admired their resilience and how they own the mission to control costs and meet demand. They are neither the most flamboyant nor do a ton of them make CEOs but I think they are the true knights of an organization – chasing parts when there are none, convincing manufacturing partners to do more, trying to outsmart suppliers to save costs, always fire-fighting; the phrase “pennies on a dollar” does not apply here.

Unless there is a system or a machine that captures EVERY SINGLE one of the supply chains’ disruptions and failures – both small and big, resolved and unresolved – and tags them within the scheme of demand realization, a truly unconstrained demand forecast is impossible.

As daunting as the task seems, it’s not! You don’t quite have to get all the way there but take baby steps in the right direction. Several organizations are already thinking about connected supply chains – a concept made possible through the IoT. Although not yet extensible to the supplier and manufacturer where most of an organization’s supply chain constraints are, it is now possible to very closely connect customer perception and supply chain performance in a business that is several times more complicated than Amazon’s.

For the first time in a long time, cost of lost sales is known factually. The benefits of doing this are quite phenomenal. But there is so much to learn and do in the space of measuring and analyzing supply chain constraints.

So, in response to all the supply chain executives that have asked me to get their demand forecast accuracy up, I just say – “Sure, but let’s get your EPS up first, shall we? Let’s start measuring the right things.”

I’d love to hear what the community of demand forecasters and planners thinks on this subject. Please do leave a comment.

(This article first appeared on LinkedIn Pulse.)

Think you’re aware of all the interesting application areas of artificial intelligence?

Find out by checking out the video below!