Chapter 2 of 4

Sizing Buffers From Forecast Error at Portfolio Scale

3 sections · ~9 min read · 3 practice questions · part of SC Analytics & Technology

Learning objectives

  • Convert a forecast-error metric into the standard deviation an inventory policy actually requires.
  • Segment a catalogue by volume and variability so that policy, not attention, is what gets assigned.
  • Allocate a fixed safety-stock budget across items to maximise demand-weighted service per dollar.

From MAD to Sigma, and From Sigma to a Buffer

Inventory formulas take a standard deviation, but forecast systems usually report mean absolute deviation. These are not the same number and substituting one for the other silently under-sizes every buffer in the catalogue. For normally distributed errors the expected absolute deviation is sigma multiplied by the square root of two over pi, which is 0.798, so sigma is MAD divided by 0.798, or about 1.25 times MAD. Skipping that conversion produces buffers roughly 20 percent too small and a service level that misses target for reasons no one can find in the policy. Two further adjustments matter as much. First, the error must be measured over the lead time, not over the forecast bucket: a weekly MAD on a four-week lead time has to be scaled by the square root of four before it means anything, and if lead time is itself variable the combined formula from Demand & Inventory Planning applies. Second, use forecast error rather than demand variability. These diverge whenever the forecast contains real information, which is the whole point of forecasting; a well-forecast seasonal item can have enormous demand variance and small forecast error, and sizing its buffer from demand variance would fund a buffer against a pattern already known in advance. The chain to hold in your head runs: measured MAD, convert to sigma, scale to the lead time, multiply by the safety factor.

Segmentation as a Policy Map, Not an Attention Map

A catalogue of ten thousand items cannot receive ten thousand individually reasoned policies, and it does not need to. Two-dimensional segmentation gives most of the benefit at a fraction of the effort. The first axis, conventionally ABC, ranks items by annual value or volume and answers how much money rides on getting the item right. The second axis, conventionally XYZ, ranks items by forecastability, usually a coefficient of variation or a MASE band, and answers how well the item can be predicted at all. The grid that results maps to policy directly. High-value, highly forecastable items justify tight review cycles and lean buffers, because the forecast can be trusted and the money is worth managing. High-value, poorly forecastable items are where buffer belongs, since no amount of modelling will fix them and their value makes stockouts expensive. Low-value, forecastable items should be automated and largely ignored. Low-value, unforecastable items are candidates for a generous buffer set once and left alone, or for delisting, because they consume planner attention out of all proportion to their contribution. The common failure is to use the grid to allocate planner attention while leaving every item on the same policy parameters. That gets the effort distribution right and the inventory distribution wrong, which is the more expensive of the two mistakes.

Allocating a Fixed Buffer Budget

Most inventory decisions in practice are not made against a service target set from first principles. They are made against a budget, and the real question is where a fixed number of dollars buys the most service. The uniform-service policy that most systems default to, a single cycle service level applied catalogue-wide, is almost always the wrong answer to that question. The cost of a unit of safety factor differs enormously across items, because it equals sigma over the lead time multiplied by unit cost and the holding rate. A high-volume, cheap, short-lead-time item can be pushed to a very high service level for very little money. A slow-moving, expensive, long-lead-time item consumes far more capital for the same nominal service level while covering a trivial share of order lines. The correct move is to buy service where it is cheap and sell it where it is expensive, then report the result as a demand-weighted service level rather than an item-weighted one. Customers experience the demand-weighted figure, since they order the fast movers far more often than the slow ones. Reporting the item-weighted average lets a catalogue look uniformly well served while the items customers actually order run short.

Worked example

Problem

Three SKUs share a safety-stock budget. SKU A: weekly forecast MAD 40 units, lead time 4 weeks, unit cost $25, annual demand 8,000 units. SKU B: weekly MAD 120 units, lead time 1 week, unit cost $8, annual demand 90,000 units. SKU C: weekly MAD 15 units, lead time 9 weeks, unit cost $60, annual demand 1,200 units. Holding rate is 25 percent per year and forecast errors are approximately normal. Compute the annual safety-stock cost of a uniform 95 percent policy (k = 1.65), then compare it with a reallocated policy that sets SKU B to 99 percent (k = 2.33), holds SKU A at 95 percent, and drops SKU C to 90 percent (k = 1.28). Report both cost and demand-weighted service.

Step by step

  1. Convert MAD to sigma: sigma = 1.25 x MAD. A = 50.0 units/week; B = 150.0; C = 18.75.
  2. Scale to the lead time: sigma_LT = sigma x sqrt(L). A = 50.0 x sqrt(4) = 100.0 units. B = 150.0 x sqrt(1) = 150.0 units. C = 18.75 x sqrt(9) = 56.25 units.
  3. Annual holding cost per unit = unit cost x 0.25. A = $6.25; B = $2.00; C = $15.00.
  4. Cost of one full unit of safety factor k = sigma_LT x holding cost per unit. A = 100.0 x $6.25 = $625. B = 150.0 x $2.00 = $300. C = 56.25 x $15.00 = $844.
  5. That ranking is the whole exercise: service on SKU B costs $300 per unit of k, while the same nominal service on SKU C costs $844, nearly three times as much, despite SKU C being by far the least variable item in absolute terms. Long lead time and high unit cost, not variability, are what make it expensive.
  6. Uniform 95 percent policy: SS_A = 1.65 x 100.0 = 165.0 units, costing 165.0 x $6.25 = $1,031. SS_B = 1.65 x 150.0 = 247.5 units, costing $495. SS_C = 1.65 x 56.25 = 92.8 units, costing $1,392. Total = $2,918 per year.
  7. Reallocated policy: SS_A unchanged at $1,031. SS_B = 2.33 x 150.0 = 349.5 units, costing $699. SS_C = 1.28 x 56.25 = 72.0 units, costing $1,080. Total = $2,810 per year.
  8. Cost change = $2,918 - $2,810 = $108 saved.
  9. Demand-weighted service under the uniform policy is 95.0 percent by construction. Under the reallocated policy it is (90,000 x 0.99 + 8,000 x 0.95 + 1,200 x 0.90) / 99,200 = (89,100 + 7,600 + 1,080) / 99,200 = 97,780 / 99,200 = 98.6 percent.

Answer. The uniform 95 percent policy costs $2,918 per year and delivers 95.0 percent demand-weighted service. The reallocated policy costs $2,810, saving $108, while delivering 98.6 percent demand-weighted service. Both cost and service improved, which is possible only because the uniform policy was buying expensive service on SKU C and declining cheap service on SKU B. Two honest limits on this result. First, the item-weighted average service fell from 95.0 to 94.7 percent, so a report built on that measure would show the reallocation as a small regression; choose the measure that reflects what customers experience before making the decision. Second, SKU C's drop to 90 percent is defensible only if nothing downstream depends on it disproportionately: a low-volume item that halts a production line is a bottleneck item and should be exempted from a purely value-weighted rule.

Practice

Work each question before opening the solution.

  1. A planning system is configured with the weekly MAD passed straight into the safety-stock formula in place of sigma, on an item with a one-week lead time. By roughly what percentage is the buffer under-sized, and what symptom would the planner see?

    Show solution for question 1

    For normal errors sigma is about 1.25 times MAD, so passing MAD directly makes the buffer 1 / 1.25 = 0.80 of what it should be, about 20 percent too small. The symptom is a service level that persistently lands below target with no visible cause: the policy documentation says 95 percent, the formula appears to be applied correctly, and measured fill rate sits several points lower on every affected item. Because it hits everything configured the same way, it looks like a systemic service problem rather than a units conversion, which is why it can survive for years.

  2. SKU C in the example has the smallest weekly MAD of the three, yet it is the most expensive item to protect. Explain why, and state which of its parameters a planner could realistically change.

    Show solution for question 2

    Cost per unit of safety factor is sigma over the lead time multiplied by unit cost and holding rate, and SKU C is punished on two of those three terms. Its nine-week lead time triples the weekly sigma of 18.75 to 56.25 through the square-root scaling, and at $60 per unit each of those units costs $15 a year to hold, against $2 for SKU B. Variability is the one term where SKU C is well behaved. Of the parameters, lead time is the realistic lever: cutting it from nine weeks to four would reduce sigma over the lead time from 56.25 to 37.5 units, a third off the buffer, with no change to forecast quality or unit cost. Unit cost is a sourcing question and the holding rate is a finance input, so neither belongs to the planner.

  3. The reallocation saved $108 a year on a $2,918 base. Why might this exercise still be worth running across a full catalogue, and what would make it not worth running?

    Show solution for question 3

    The dollar saving is small because the example has three items, but the mechanism scales while the effort does not: the same rule applied across ten thousand SKUs is a query and a parameter update, and the service gain of 3.6 demand-weighted points is the larger prize, since it lands on the fast movers customers order most. What makes it not worth running is a catalogue where cost per unit of safety factor is roughly uniform across items, because then there is nothing to arbitrage and a uniform policy is already close to optimal. Check the spread of that statistic first; if the highest and lowest items are within a factor of two, the reallocation will not repay the effort of implementing and explaining it.