“Trillions of records of millions of people…Finding the useful and right information, understanding its quality and producing reliable analyzed data in a timely and cost-effective manner are all critical issues.”

Smart Software Senior Vice President for Research Tom Willemain recently had the opportunity to talk with Dr. Mohsen Hamoudia, President of the International Institute of Forecasters (IIF), to discuss current issues with, and opportunities for, big data analytics. The IIF informs practitioners on trends and research developments in forecasting via print and online publications and the hosting of professional conferences.

Dr. Hamoudia begins, by way of introduction:

In all industries, data availability is exploding in volume, variety and velocity. Big data analytics is playing an important role in identifying the data that is most important to the business.

Let me take the example of the Information and Communication Technology (ICT) sector. We are seeing literally exponential growth in the amount of data available to telecoms, Over-the-top (OTT) independent content distributors, government, regulators and other organizations.

Around the world, we are witnessing petabytes of data: trillions of records of millions of people—all coming from multiple sources. Among these sources: internet connections, sales, customer contact centers, social media, mobile and land lines data. Finding the useful and right information, understanding its quality and producing reliable analyzed data in a timely and cost-effective manner are all critical issues. ICT companies are increasingly looking to find actionable insights in their data. How they can increase their customer base and loyalty programs? How can they improve the quality of service (QoS) and reduce customer turnover? With the right big data analytics platforms in place, they can be more competitive and efficient, improving operations, customer service and risk management. Forecasting and predicting customer trends and directions are key for telecoms.

Forecasting skills, including mathematics, statistics and econometrics, form one of the most important “blocks” of required skills in managing Big Data. Some forecasting activities naturally form part of the big data debate.

In retail industries, forecasting addresses daily demand across thousands of products. Financial forecasting, whether considering customer behavior or financial data series, generates massive on-line data sets. As pointed out by Robert Fildes, Distinguished Professor at Lancaster University, as yet the academic forecasting community is not thoroughly engaged—with only a few exceptions. Hal Varian of Google has looked at some of the work that David Hendry and Jennifer Castle, at Oxford University, have undertaken on searching large data sets for data congruent meaningful models. Stock and Watson have also developed their own approaches to large macro data sets. But despite the attempt, at last year’s symposium on forecasting in Seoul, to explore the theme of big data and its forecasting applications, there remain few convincing applications of using on-line data on real forecasting problems.

Q. One hears a great deal about “predictive analytics” these days, yet the phrase rarely is linked with forecasting. Do you agree that forecasting lies at the heart of predictive analytics? Have you an explanation for why the link has been broken? Have you ideas about how to re-inject forecasting into the conversation?

The results of forecasting (the “what”) are perhaps now perceived as less important than the “how”. Consequently, the trust that users give to traditional forecasting has declined. Who indeed is challenging the accuracy or relevance of forecasting by comparing, a posteriori, the reality vs. forecast—making a case for metholodiges’ effectiveness and therefor building credibility?

With the current perception of “predictive analytics”, there is probably more space in the public imagination allocated to the “how” side of things, and therefor a more credible story to tell to partners, investors or customers.

Q. It appears that there is almost no link between traditional forecasting and mobile technology (smart phones, tablet computers). Is this true, or are some companies migrating forecasting to mobile devices? Do you see a path forward in which traditional forecasting algorithms would routinely reside on mobile devices?

First of all, I am really delighted to invite your readers to have a look at our latest issue of Foresight. An excellent paper on the subject, “Forecasting In the Pocket: Mobile Devices Can Improve Collaboration”, explains that “the increasing popularity of PDAs, smartphones, tablet computers and other mobile devices opens up new opportunities for communication and collaboration on business forecasts.” The authors tell us “mobile forecasting (m-forecasting) applications may streamline approaches to collaboration between retailers and suppliers, thus contributing to the provision and exchange of product information, especially since forecasts are strongly tied to local context knowledge.”

For example, on the ICT & OTT side, a large number of predictive projects, such as those of Google+ and Facebook, are happening thanks to the inclusion of the “user location” data in the OTT IT systems. In my opinion, and what I see in some sectors like retail and logistics, is that traditional forecasting and mobile forecasting (m-forecasting) are complementary. This latter could be seen as a bottom-up forecasting approach that will or will not confirm the top-down forecasting results.

Q. Some people argue that big data will facilitate the replacement of forecasting with “sense and react” systems. Practically speaking, how would you explain “sense and react”, and are there application areas where you think it is or is not likely to take hold?

It seems to me that “sense and react” is fully oriented to the short-term perspective. Forecasting extends this by addressing needs for a variable horizon: short-term and medium-term.

As a side effect of ATAWAD (Anytime, Anywhere, Any Device), the decision-making criteria are, more than ever, “short term”. Big data is a “weak signals” sensing system, which enable the near-real-time detection of business opportunities that would go unnoticed with traditional IT systems. There are not really preferred or non-priority applications for this, the question is more on the “when” side.

Big data is relevant when looking below the surface in difficult economic times, but I am less sure whether it is worth the effort in “normal” economic period. To conclude on this point: I will be happy to see an example on how accurate are forecasts which are based on “sense and react” versus forecast based on traditional models.

Q. I’m asking some big questions. To what extent do you see the IIF community shaping these discussions and outcomes? How can readers join in the dialogue?

We are expecting an increasing availability, and increasing usage, of huge amount of data in many industries—such as energy, transportation, health care, finance, telecommunications and tourism.

Many of the IIF’s members are engaged in different aspects of the big data “movement.” The IIF is doing some work in the forecasting activities that naturally form part of the big data debate. More generally, the IIF is actively participating in, and providing a forum for, the discussion of forecasting in the wider world.

The theme of our last International Symposium on Forecasting (ISF) held in Seoul was “Forecasting with Big Data” and a few presentations were related to health care and telecommunications. A relevant workshop has just been run by the European Central Bank (ECB). If these models are capitalized on, they have the potential to impact the economic policy of Europe quite quickly.

Readers can join in the dialogue by contributing papers to the IIF’s publications (The International Journal of Forecasting, Foresight and The Oracle). Foresight, for one, is an invaluable voice in bringing academics and practitioners together in an ongoing discussion.

Readers also can present papers at the annual conference (the aforementioned ISF). They also can suggest and organize specific workshops for specific applications of big data, like the one that was just organized by the ECB in Frankfurt. Another opportunity is to invite IIF’s members to attend any meeting related to forecasting with big data. All these opportunities form good platforms for networking and working together.

Mohsen Hamoudia, PhD, is the President of the International Institute of Forecasters. He also serves as Head of Strategy for Large Projects (Paris) for Orange Business Services (the former France Telecom).

Thomas Willemain, PhD, co-founded Smart Software and currently serves as Senior Vice President for Research. Dr. Willemain also serves as Professor Emeritus of Industrial and Systems Engineering at Rensselear Polytechnic Institute, and as a member of the research staff at the Center for Computing Sciences, Institute for Defense Analyses.

Demand planning takes time and effort. It’s worth the effort to the extent that it actually helps you make what you need when you need it.

But the job can be done well or poorly. We see many manufacturers stopping at the first level when they could easily go to the second level. And with a little more effort, they could go all the way to the third level, utilizing probabilistic modeling to convert demand planning results into an inventory optimization process.

The First Level

The first level is making a demand forecast using statistical methods. Figure 1 shows a first level effort: an item’s demand history (red line) and its expected 12-month forecast (green line).

The forecast is bare bones. It only projects expected demand ignoring that demand is volatile and will inevitably create forecast error. (This is another example of an important maxim: “The Average is Not the Answer”). The forecast is as likely to be too high as it is to be too low, and there is no indication of forecast uncertainty accompanying the forecast. This means the planner has no estimate of the risk associated with committing to the forecast. Still, this forecast does provide a rational basis for production planning, personal scheduling, and raw materials purchase. So, it’s much better than guessing.

The Second Level

The second level takes explicit account of forecast uncertainty. Figure 2 shows a second level effort, known as a “percentile forecast”.

Now we see an explicit indication of forecast uncertainty. The cyan line above the green forecast line represents the projected 90th percentile of monthly demand. That is, the demand in each future month has a 90% chance of falling at or below the cyan line. Put another way, there is a 10% chance of demand exceeding the cyan line in each month.

This analysis is much more useful because it supports risk management. If it is important to assure sufficient supply of this item, then it makes sense to produce to the 90th percentile instead of to the expected forecast. After all, it’s a coin flip as to whether the expected forecast will result in enough production to meet monthly demand. This second level forecast is, in effect, a rough substitute for a careful inventory management process.

Figure 2. A percentile forecast, where the cyan line estimates the 90th percentiles of monthly demand.

Going All the Way to the Third Level

Best practice is the Third Level, which uses demand planning as a foundation for completing a second task: explicit inventory optimization. Figure 3 shows the fundamental plot for the efficient management of our finished good, assuming it has a 1 month production lead time.

Figure 3 shows the utilization of probabilistic forecasting and how much draw-down in finished good inventory might take place over a one month production lead time. The uncertainty in demand is apparent in the spread of the possible demand, from a low of 0 to a high of 35, with 15 units being the most likely value. The vertical red line at 22 indicates the “reorder point“ (or “min” or “trigger value”) corresponding to keeping the chance of stocking out while waiting for replenishment to a low 5%. When inventory drops to 22 or below, it is time to order more. The Third Level uses probabilistic demand forecasting with full exposure of forecast uncertainty to efficiently manage the stock of the finished product.

To Sum Up

Forecasting the most likely demand for an item is a useful first step. It gets you halfway to where you want to be. But it provides an incomplete guide to planning because it ignores demand volatility and the forecast uncertainty that it creates. Adding a cushion to the demand forecast gets you further along, because it lessen the risk that a jump in demand will leave you short of product. This cushion can be calculated by probabilistic forecasting approaches that forecasts a high percentile of the distribution of future demand. And if you want to take one step further, you can feed forecasts of the demand distribution over a lead time to calculate reorder points (mins) to ensure that you have an acceptably low level of stock-out risk.

Given what modern forecasting technology can do for you, why would you want to stop halfway to your goal?

Managing inventory requires executives to balance competing goals: high product availability versus low investment in inventory. Executives strike this balance by stating availability targets and budget constraints. Then supply chain professionals translate these “commander’s intentions” into detailed specifications about reorder points and order quantities.

A High-Stakes Race Between Supply and Demand

Let’s focus on reorder points (also known as mins). They work as follows. As on-hand inventory decreases in response to demand, it eventually drops down to or below a trigger value, the reorder point or min. At that point, it’s like a gun goes off to start a race between supply and demand. A replenishment order is sent to restock the item, but there is a replenishment lead time, so the restocking is not instantaneous. While your system waits for resupply, demand continues to whittle away at the stock on hand. It is bad news if demand wins the race, because then you won’t be in position to provide what somebody is demanding. Then they either get it from a competitor or get back-ordered and unhappy: either way, stocking out is a bad outcome for you and your customer.

The risk of stocking is controlled by your staff’s choice of reorder points. If they are set too high, stock-outs are rare but inventory is bloated. Set them too low and stock-outs abound. So how should reorder points be set?

Avoiding Foolish Follow-Through

Several factors govern stock-out risk. Each item in your inventory has its own demand history and lead time. Together with your chosen availability targets, these factors determine the best choice of reorder point. But the relationships are statistical and require good analysis to work out. Inventory Optimization Software can compute the proper reorder point for each of tens of thousands of items. But instead of relying on proper analysis, many companies fall back on simple rules of thumb or just “doing what we always do”.

In place of using the right math, companies often rely on rules of thumb that serve them poorly. Here are some examples in order of most common to least common.

1) Multiples of Average Demand

Setting reorder points at some (arbitrary) multiple of average demand starts to rely on actual facts. But it ignores the key demand attribute that drives stock-out risk: demand variability. Two items with the same average demand but very different levels of variability will require very different reorder points to insure the same low risk of stock-out. (See Figure 1)

2) Gut feel

Some companies have self-styled supply chain gurus. Even if they actually are Jedi masters, it’s impossible to keep up with tens of thousands of items whose reorder points should be reviewed frequently.  And if the logic that drives decision making is buried in a hard to use spreadsheet that only they know how to use, the company risks not being able to execute the inventory plan without that one individual –a risky proposition.

3) Average Demand + some multiple of Demand Variability

This approach is taught in many “Inventory 101” courses. But it implicitly assumes some facts about demand that are very often not true: that demand has a Normal (“bell-shaped”) distribution and that demand in one period does not relate to demand in the previous time period(s).  Assumptions of independence and reliance on normal distribution models just don’t cut it.

4) Nursery rhymes

Not at all the norm, hence being last on the list, but we heard of one company that used one simple rule for all items: “If it’s down to four, order more”. It’s crazy to believe that one rule applies to all items at all times. But at least it rhymes.

Your people can do better than to rely on any of these approaches. Do you know whether your company is using any one of them?

Getting It Right

The right way to set reorder points uses the tools of probability theory. The details depend on whether you are selling finished goods or spare parts. Spare parts are usually more difficult to manage because they have quirky demand patterns: high intermittency (lots of zero demands), high skewness (lots of small demands but with some whoppers too), and auto-correlation (“feast or famine” behavior). Modern Reorder Point Software takes these quirks into account to set reorder points that insure the desired level of item availability. Importantly, they also let your people see explicit trade-off curves, so they can strike the balance you want — at the item by location level – between stock-out risk and inventory investment.

Inventory is a major item on the balance sheet and needs high-level attention. At many manufacturers, service parts can represent up to half of revenue. Modern software lets the C-Suite move beyond, incomplete math and other inadequate approaches to managing inventory.

Figure 1:  Two equally important items with the same average demand get assigned the same stocking policy that determines the Min (reorder point) as 2 x average lead time demand.  Despite the “same” stocking policy service performance varies significantly with the stable Item A experiencing overstocks and the volatile Item B experiencing stock outs.

In our travels around the industrial scene, we notice that many companies pay more attention to inventory Turns than they should. We would like to deflect some of this attention to more consequential performance metrics.

Recall the definition: Turns = Annual dollar cost of goods sold / Average dollar value of inventory. If you sell $1 million of stuff in a year and have an average of $100,000 of stuff on the shelf each day, you are running at an impressive 10 Turns (Walmart runs at around 8). Supposedly, having high Turns signals efficient management, and keeping your Turns higher than competitors’ signals competitive advantage.

But as happens with most performance metrics, there is more to the story. Turns may be very salient to the CFO, but they can be a straightjacket to the COO. This is because Turns are not directly related to customer service; in fact, high Turns can be synonymous with low service levels and fill rates. S&OP consultant Darrin Oliver calls Turns his “pet peeve metric” because “the customer doesn’t care about Turns.”

Suppose you are unhappy with your current Turns value. What can you do to boost the number? Since Turns is a ratio, you can increase it by either increasing the numerator (goods sold) or decreasing the denominator (inventory). Increasing sales is more difficult because it requires the cooperation of the customer. Decreasing inventory is easier because it’s completely under your control: just make smaller replenishment orders, which also saves money in the short run. Indeed, you can get very enthusiastic and cut inventory to the bone. You end up with a better looking number for Turns—and a serious problem with stockouts, backorders, lost sales, lost customer good will and lost market share. Who’s sorry now?

Here’s a lightly edited version of a story on this topic told by a very wise practitioner. “Back in my other life they were all about improving Turns. Why, I have no idea. So I pointed out the risks that you run. And they really weren’t interested. So we took our global inventories down to [a lower level], and then were breaking on stock left and right on a daily basis. Our turns were great, but we weren’t making any money, because we couldn’t get anything out the door, because we didn’t own it. The higher your turns, the lower your inventory’s going to have to be, or you’re just going to have really good flow. And in our industry that’s a very, very difficult thing to achieve. So if we can have reasonable Turns but still be in stock, I think that’s what we want to do. It can be very frustrating in an operations world to try to explain what we do every day and what the risks to the business are when the financial people are just looking at one or two metrics. They’re basically trying to plan the business in a vacuum, and it’s very difficult and very risky to do that.”

Thomas Willemain, PhD, co-founded Smart Software and currently serves as Senior Vice President for Research. Dr. Willemain also serves as Professor Emeritus of Industrial and Systems Engineering at Rensselear Polytechnic Institute and as a member of the research staff at the Center for Computing Sciences, Institute for Defense Analyses.

Demand planners have to cope with multiple problems to get their job done. One is the Irritation of Intermittency. The “now you see it, now you don’t” character of intermittent demand, with its heavy mix of zero values, forces the use of advanced statistical methods, such as Smart Software’s patented Markov Bootstrap algorithm. But even within the dark realm of intermittent demand, there are degrees of difficulty: planners must further cope with the potentially costly Scourge of Skewness.

Skewness is a statistical term describing the degree to which a demand distribution is not symmetrical. The classic (and largely mythic) “bell-shaped” curve is symmetric, with equal chances of demand in any time period falling below or above the average. In contrast, a skewed distribution is lopsided, with most values falling either above or below the average. In most cases, demand data are positively skewed, with a long tail of values extending toward the higher end of the demand scale.

Bar graphs of two time series
Figure 1: Two intermittent demand series with different levels of skewness
Figure 1 shows two time series of 60 months of intermittent demand. Both are positively skewed, but the data in the bottom panel are more skewed. Both series have nearly the same average demand, but the one on top is a mix of 0’s, 1’s and 2’s, while the one on the bottom is a mix of 0’s, 1’s and 4’s.

What makes positive skewness a problem is that it reduces an item’s fill rate. Fill rate is an important inventory management performance metric. It measures the percentage of demand that is satisfied immediately from on-hand inventory. Any backorders or lost sales reduce the fill rate (besides squandering customer good will).

Fill rate is a companion to the other key performance metric: Service level. Service level measures the chance that an item will stock out during the replenishment lead time. Lead time is measured from the moment when inventory drops to or below an item’s reorder point, triggering a replenishment order, until the arrival of the replacement inventory.

Inventory management software, such as Smart Software’s SmartForecasts, can analyze demand patterns to calculate the reorder point required to achieve a specified service level target. To hit a 95% service level for the item in the top panel of Figure 1, assuming a lead time of 1 month, the required reorder point is 3; for the bottom item, the reorder point is 1. (The first reorder point is 3 to allow for the distinct possibility that future demand values will exceed the largest values, 2, observed so far. In fact, values as large as 8 are possible.) See Figure 2.

Histograms of two time series
Figure 2: Distributions of total demand during a replenishment lead time of 1 month
(Figure 2 plots the predicted distribution of demand over the lead time. The green bars represent the probability that any particular level of demand will materialize.)

Using the required reorder point of 3 units, the fill rate for the less skewed item is a healthy 93%. However, the fill rate for the more skewed item is a troubling 44%, even though this item too achieves a service level of 95%. This is the scourge of skewness.

The explanation for the difference in fill rates is the degree of skewness. The reorder point for the more skewed item is 1 unit. Having 1 unit on hand at the start of the lead time will be sufficient to handle 95% of the demands arriving during a 1 month lead time. However, the monthly demand could reach above 15 units, so when the more skewed unit stocks out, it will “stock out big time”, losing a much larger number of units.

Most demand planners would be proud to achieve a 95% service level and a 93% fill rate. Most would be troubled, and puzzled, by achieving the 95% service level but only a 44% fill rate. This partial failure would not be their fault: it can be traced directly to the nasty skewness in the distribution of monthly demand values.

There is no painless fix to this problem. The only way to boost the fill rate in this situation is to raise the service level target, which will in turn boost the reorder point, which finally will reduce both the frequency of stockouts and their size whenever they occur. In this example, raising the reorder point from 1 unit to 3 units will achieve a 99% service level and boost fill rate to a respectable, but not outstanding, 84%. This improvement would come at the cost of essentially tripling the dollars tied up in managing this more skewed item.

Thomas Willemain, PhD, co-founded Smart Software and currently serves as Senior Vice President for Research. Dr. Willemain also serves as Professor Emeritus of Industrial and Systems Engineering at Rensselear Polytechnic Institute and as a member of the research staff at the Center for Computing Sciences, Institute for Defense Analyses.