Forecasting and Inventory Planning for Parts with Intermittent
Demand - A Case Study
Gokul Hariram Kottaram Venkitachalam, David B. Pratt, and Camille F. DeYong
School of Industrial Engineering and Management
Oklahoma State University
Stillwater, OK 74078, USA
Steven Morris and Michel Leonard Goldstein
School of Electrical and Computer Engineering
Oklahoma State University
Stillwater, Ok., 74078, USA
Abstract
Forecasting demand and developing inventory strategies for parts with an intermittent demand history presents a
formidable challenge. We review the results of efforts to forecast part failures and determine inventory strategies for
military aircraft parts. Applications of bootstrapping and Croston's method are summarized and the results
contrasted with more traditional time series approaches.
Keywords
Intermittent Demand, Forecasting, Inventory Planning.
1. Intermittent Time Series
An intermittent time series is a time series of non-negative integers where some of the values are zero [1]. This paper
explores forecasting and inventory planning of intermittent series characterized by a time series pattern that contains
frequent and irregularly spaced zero values. Table 1 shows an example of a time series that exhibits this property.
Time series of this type are typical in the early failure histories of aircraft repair parts.
Table 1. Intermittent Data Example
Time Period 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Demand 2 1 5 0 0 0 0 1 0 0 0 0 1 0 1
2. Traditional Techniques
When traditional forecasting techniques are applied to an intermittent series, the results are frequently
unsatisfactory. As a typical example, consider the application of a 3-period moving average model with one period
lead-time to the data in Table 1. The resulting forecasts and absolute forecast errors are shown in Table 2. The mean
absolute error over the twelve time periods (period 4 to period 15) is 0.86. Contrast this with the naïve strategy of
forecasting zero for each time period. This naïve strategy generates a mean absolute error over the twelve time
periods of 0.25. This raises the question of whether there is any benefit in attempting to forecast at all.
Table 2. 3-Period Moving Average Forecast and Error
Time Period 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Demand 2 1 5 0 0 0 0 1 0 0 0 0 1 0 1
Forecast 2.7 2.0 1.7 0.0 0.0 0.3 0.3 0.3 0.0 0.0 0.3 0.3
Abs. Error 2.7 2.0 1.7 0.0 1.0 0.3 0.3 0.3 0.0 1.0 0.3 0.7
Table 3 shows the results of applying a simple exponential smoothing model to the example data. For this model, the
parameter alpha is set to 0.2 and the initial forecast is set equal to the initial demand. The mean absolute error over
the twelve time periods is 0.99, which is once again inferior to the naïve approach of forecasting zeroes.
It can legitimately be argued that the viability of traditional forecasting approaches is sensitive to the
"intermittenness" of the series being forecast and the parameters selected for the model. However, the results shown
above are typical and representative of those that were achieved in the case study presented in this paper.
Table 3. Simple Exponential Smoothing Forecast (alpha=0.10) and Error
Time Period 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Demand 2 1 5 0 0 0 0 1 0 0 0 0 1 0 1
Forecast 2.0 2.0 1.8 2.4 2.0 1.6 1.2 1.0 1.0 0.8 0.6 0.5 0.4 0.5 0.4
Abs. Error 2.4 2.0 1.6 1.2 0.0 1.0 0.8 0.6 0.5 0.6 0.5 0.6
3. The Case Under Study
The time series under consideration in this study are order and delivery data covering 163 periods for approximately
30,000 parts that are uniquely used on the KC-135 [2]. The ongoing project of which this study is a component aims
to improve the forecasting of demand for these (and other) parts that have very low demand. Of particular interest
are parts that experience demand “surges”. Surge parts are defined as parts that have little or no demand over the
history of the aircraft and then suddenly experience a large increase in demand [2]. Each part under consideration
can be associated with a unique identification number known as an NIIN. Croston’s method and a bootstrapping
technique are used in conjunction with a clustering technique in an attempt to improve the forecast of demand for
these intermittent parts.
3. Inventory Aspects
An important aspect of the problem being considered is to understand the inventory implications of the forecasts.
Failed parts (a "demand" in this study) are assumed to be non-repairable and must be replaced from inventory or
ordered. If demand for a part is incorrectly forecast, then either an inventory excess or an inventory shortage will
occur. If the part is critical to the safe and effective operation of an aircraft, then a shortage may well lead to a
grounded aircraft. This aircraft could potentially be a fighter jet (not the example considered here) or a critical
supplies transport, in which case, the strategic importance of the loss of capability incurred cannot easily be
quantified in dollars. Thus, it is of paramount importance, where possible and economically viable, to have the
fewest number of aircraft grounded. The implication of this aspect of the problem is that the impact of inventory
shortages and inventory excesses may be very different. A shortage in inventory may be more harmful than an
excess in inventory. The quantification of this difference is elusive and is handled in this study through the use of a
sensitivity analysis. The sensitivity analysis explores a range of ratios (cost of shortage/cost of excess) to determine
if this factor is significant in determining the preferred forecasting approach.
A second significant inventory aspect of this problem is the consideration of lead-time. The majority of the parts
considered in this study have an average order lead-time of eight periods. That is, an order place in period 1 will
arrive and be available for use in period 9. This long lead-time magnifies the impact of shortages and the impact of a
forecasting model that is slow to respond to changes in the underlying demand pattern.
4. Measure of Error
Some forecast analysts believe that error measures should account for asymmetries in the cost of errors [3]. In other
words, a positive forecast error is treated differently than a negative forecast error. Traditional measures of error,
such as MSE (mean squared error), fail to effectively evaluate forecasting methods when the costs of errors are
asymmetrical. Since the present case study presents a situation where asymmetry is important, this necessitates the
development of a forecast evaluation method that considers the differing levels of impact due to shortages and
excesses in inventory. This evaluation is handled through a sensitivity analysis. The sensitivity analysis focuses on
an average error measure without regard to a measure of variance in the error terms. Consideration of variance is
under consideration for a later phase of the project. The first step in the analysis is to compute the error in each
period generated from using a particular forecasting approach. On-hand inventories and order receipts for each
period are accounted for in these calculations. In each period, either a shortage or an excess will occur. Tallies are
kept of the cumulative number part shortages and part excesses. Summing the total number of shortages and the
total number of exc esses across the forecasted time periods determines the initial performance measure of the
forecasting method. Table 4 illustrates this calculation. This calculation is repeated for each part and each
forecasting methods under consideration.
Table 4.Calculation of Total Shortages and Excesses
Time Period 0 1 2 3 4 5 6 7 8 9 10 11
Inventory 0 18
Fcst Error -2 -1 -5 0 0 0 0 -1 0 0 0
Shortages 2 3 8 8 8 8 8 9 9 Shortages 63
Excesses 9 10 Excesses 19
A sensitivity analysis is then conducted to determine the sensitivity of selecting a preferred forecasting method to
the increasing weight (importance) placed on shortages. This is accomplished by multiplying the number of
shortages by a weighting factor prior to calculating the final sum. A hypothetical example of this sensitivity analysis
is illustrated in Table 5. Table 5 is then used to determine which method dominates at each level of the sensitivity
analysis. Dominance is established by the forecasting method that generates the lowest value of the weight sum of
shortages and excesses. The highlighting in Table 5 illustrates identification of dominance. A forecasting method
may dominate across all the assigned weights. Alternatively a forecasting method can dominate only beyond a
threshold weight (as shown in Figure 5) or dominance can shift between the forecasting methods with no clear
pattern.
Table 5. Sensitivity analysis and determination of the dominant forecasting technique
Part #1 Part Period Increasing Weight on Shortages
Forecasting
Method
Shortage
s
(S)
Excesses
(E) S+E 1.5S+E 2S+E 5S+E 7S+E 10S+E 20S+E 50S+E
Method 1 40 44 84 104 124 244 324 444 644 844
Method 2 44 106 150 172 194 326 414 546 766 986
Method 3 28 98 126 140 154 238 294 378 518 658
5. Croston's Method
Croston’s method deals with the problem of forecasting demand levels when the demand patterns are not regular [4].
The method was proposed in 1972 and has since established itself as the standard approach to forecasting problems
with irregular patterns. Croston observed that the use of the traditional exponential smoothing for intermittent
demands is not suitable, since it tends to overestimate the demand levels. The alternative that Croston proposed
involves breaking the intermittent demand time series into two constituent time series - one series for the non-zero
demand values and other series for the time interval between the non-zero demand values. Traditional exponential
smoothing is then used on each of the constituent parts separately. As an illustration, Table 6 shows the two
Croston’s constituent time series that would result from the data in Table 1.
Table 6. Croston's Constit