Understanding Intermittent Demand Forecasting Solutions

Charles Smart

Charles Smart

When a customer calls for that product item that almost no one ever asks for, do you have it?  If your answer is no, it could cost you a sale, or even a customer.  Equally important, if that item is only occasionally requested, do you have too many units of the product on-hand to avoid stocking out of it?  If your answer is yes, then excess inventory is probably costing you money.

Forecasting intermittently demanded, “slow-moving” product items or parts is a problem that is especially well-known to managers in service parts organizations and companies in the capital goods industries, among others.  There are two common mistakes these firms make when forecasting intermittent demand:  First, they focus on estimates of per period demand when they should really focus on estimates of the inventory stocking requirements necessary to meet their desired service levels.  Second, they use forecasting methods that are inappropriate for intermittent demand.

Not all forecasting methods are created equal

Traditional forecasting methods, such as exponential smoothing and moving averages, that are designed for normal, high-volume demand just don’t work well with intermittent demand.  Even worse, methods that some companies use specifically for intermittent demand, such as Croston’s method and Poisson models, fail to provide accurate inventory stocking recommendations.  And judgmental, “gut feel” techniques used by some organizations are neither feasible nor adequate, especially when you are trying to forecast thousands of items at a time.

Here’s the problem:  Traditional forecasting methods, commonly used by ERP/SCM and other forecasting software systems, fail because they try to identify recognizable patterns in the demand data, such as trend and seasonality.  However, intermittent demand data don’t exhibit such regular patterns and tend to be characterized by a preponderance of zero values.

The older, traditional technologies ignore the special role of zero values when analyzing demand and tend to simply “smooth over” the zero and occasional non-zero values found in intermittent data.  But the zeroes are very important!  If you don’t properly account for the timing and frequency of zero values, you can’t generate an accurate demand forecast or properly estimate the required inventory level for an intermittent item.  However, new  empirical methods like statistical “bootstrapping” don’t make assumptions about patterns in the data, do account for the zero values found in intermittent demand, and can produce accurate forecasts and inventory stocking estimates.

Figure 1 below shows the challenge demand planners face. It plots the demand over 36 months for three intermittent part items (shown in red, blue and green),.  Many months had no demand at all (zero values), and when demand did appear, its value varied erratically.

Figure 1

Figure 1

Figure 1: Examples of Monthly Demand for Intermittently Demanded Parts

Evaluating intermittent demand forecasting software

When you evaluate forecasting software, it’s important to “look under the hood”, because how the software does its job is just as important as what it does.  Of course, most companies need software that handles a variety of forecasting situations, including forecasting high volume, frequently demanded products.  But, software that solves the intermittent demand forecasting problem is very specialized and should provide certain capabilities that enable it to detect intermittent demand patterns and create accurate forecasts which facilitate demand planning and cost-effective inventory management decisions.

When evaluating software for intermittent demand forecasting, here are three things you should look for:

  1. The forecasting method should accurately forecast the entire distribution of lead time demand values (i.e., total demand over a lead time), rather than produce just a single number representing the average demand per period.
  2. The solution should provide optimal service level inventory requirements for satisfying total lead time demand (for example, the minimum safety stock and inventory needed for a 90%, 95% or 99% likelihood of not stocking out of a product item).
  3. The solution should accurately reflect the asymmetrical (i.e., non-normal) nature of the lead time demand distribution—a phenomenon typical of intermittently demanded items, as illustrated in the demand distribution in Figure 2 below.

Figure 2

Figure 2

Figure 2: Distribution of Lead Time Demand for a Service Part Exhibiting Intermittent Demand

In addition, to make sure that you find a comprehensive demand forecasting solution that works for all of your company’s needs, you should:

  1. Investigate and understand the system’s capabilities to satisfy other needs you may have besides intermittent demand forecasting, such as promotion/event modeling, new product forecasts, multi-level forecasting at the item and group level, multivariate regression for causal analysis, and automatic forecasting of large-scale forecasting jobs, among other needs;
  2. Educate yourself about forecasting software options at events such as those sponsored by the Institute of Business Forecasting;
  3. Test drive solutions to compare results and confirm that you can achieve the accuracy and value you expect.  In particular, make sure that the inventory recommendations provided by a possible solution actually hit the service level targets you specify; this ensures that future decisions you make about your inventory investment will be the correct ones.

By making the right forecasting software decision, the payoff can be tremendous.  Depending on the size of your company’s inventory, an accurate solution for forecasting intermittent demand can lead to first-year inventory savings in the millions of dollars and service level improvements of 10% to 20% or more.  These kinds of results free up cash for other uses, reduce the number of lost sales, help to guarantee better customer service, and increase shareholder value.

Charles Smart
President and CEO
Smart Software, Inc.

See CHARLES SMART at The IBF’s:

$695 (USD) for 3 Full Days!

October 12-14, 2009
Orlando Florida USA

14 Responses to Understanding Intermittent Demand Forecasting Solutions

  1. One thing to bare in mind – causal forecasting. i.e. let’s not only look at distributions over time but also consider the macro factors such as promotions, competitors etc.

  2. Why do you suggest that corston’s and other intermittent forecasting techniques are less that adequate? What is your alternative?

    Thank you,
    Jim Burns
    Executive Director, Sales Forecasting
    Warner Home Video, Inc.
    Office: 818.977.6824
    Mobile: 310.648.4973
    Jim.Burns @ warnerbros.com

  3. Jim,
    Intermittent Demand is tough – no doubt about it.

    Can’t speak for all of them, but Croston’s method (as well as others) tend to introduce some bias as a result of the nature of demand.
    ( http://eprints.lancs.ac.uk/7060/)

    Smart’s is actually pretty good in the real world scenarios I’ve seen so far. Hope that helps.
    Mike

  4. Jim:

    Thanks for your response to my blog.

    Croston’s method and some of the other dedicated intermittent demand forecasting (IDF) techniques are OK at projecting a single number representing the average demand per period. Where they tend to have problems is in generating an accurate distribution of all possible values of total cumulative demand over a product’s lead time. Croston’e method, in particular, assumes that this distribution will be normal, i.e., look like a bell-shaped curved. In fact, lead time demand distributions for intermittently demanded items tend to be asymmetric with long tails, like the distribution shown in Figure 2 in the blog.

    In our demand forecasting and planning system, SmartForecasts, we offer a patented IDF solution (based on NSF-funded research) that is empirically based and uses statistical bootstrapping technology to generate the lead time demand distribution. From this, we can calculate optimal safety stock and inventory requiremments for any desired service level, in addition to the average demand per period. If you would like to learn more about our approach, I would invite you to visit our web site at http://www.smartcorp.com/intermittent_demand_planning.asp.

    Regards,

    Charles

    Charles N. Smart
    President & CEO
    Smart Software, Inc
    Ph: (800) 762-7899 or (617) 489-2743
    Email: CharlesS@smartcorp.com
    Web site: http://www.smartcorp.com

  5. Jim:

    There are a variety of different demand patterns which can all be classified as intermittent to some extent. There is the classic Croston’s model, which corresponds to having a single customer who orders about the same quantity every few months or so. There are product life cycle models which correspond to cases where old products are superceded by new ones resulting in only intermittent demand for the old products. There is also seasonal intermittent demand for products like Halloween candy, and many other different cases. I don’t think there is a single model which handles all of these situations effectively – you just need to have a Swiss Army Knife in your stats toolbox to be able to model each situation appropriately.

    Regards,

    Rudi Pizzano
    RoadMap Technologies
    rudi@roadmap-tech.com

  6. Hi, Charles,
    I read some articles stating the poisson distribution also can handle the slow moving item and intermittent demand item. What do you think?
    By the way, in what way you can identify which products have the intermittent demand pattern and /or slow moving? And how many periods’ demand data should be used for analysis at the minimum in time bucket, weeks, months?
    please advise,
    jon

  7. Jon:

    Thanks for your comment and sorry for the delay in getting back to you.

    Yes, I have heard of some people attempting to use a variant of the Poisson model to handle intermittent demand. Howvever, we feel that the empiricaly-based statistical bootstrapping solution that we have patented and incorporated in our SmartForecasts system gives consistently accurate estimates not only of the demand forecasts but also of the corresponding safety stock and inventory requirements for an intermittently demanded item.

    There are different ways to define intermittent demand, but practically speaking, intermittently demanded items normally have many periods of zero demand interspersed with seemingly random-occuring spikes of non-zero demand. If your demand data has this pattern and the proportion of periods with zero demand is at least 25-30% (and in certain cases, like the demand for service/spare parts, often much greater than this), you probably have intermittently demanded items.

    If you have any further questions about intermittent demand or our approach to forecasting it, please feel free to contact me directly at charless@smartcorp.com.

    Regards,

    Charles Smart
    Smart Software, Inc.

  8. HI Charles,
    Very good article,
    May I ask how many periods at minimum to make bootstrapping method to work?
    In Daily time bucket, weekly or monthly time buckets?
    If in daily time bucket, how many days at minimum are required?
    If weekly, how many weeks at minimum are required?
    If monthly, how many months at minimum are required?
    In my case, the lead time is 6 working days for replenishment for most parts? some parts have 3 working days.
    please advise,
    thanks

  9. Hi Jon,

    I’m answering your question above on behalf of Charles here.

    We typically request 1 -2 years of actual demand. The more granular the better the results will be. However, weekly or monthly data works well for many companies. Feel free to contact me with additional questions at gregh@smartcorp.com

    Thanks,
    Greg Hartunian
    President and CEO
    Smart Software

Leave a Reply

Your email address will not be published. Required fields are marked *

WordPress Anti Spam by WP-SpamShield