MBA Bullwhip Effect Concept Card Template

Bullwhip Effect Concept Card: [Your course or supply chain]

A concept card for the supply chain elective: why order variability grows as orders move upstream from the shelf to the factory, how to measure it, how much of it a forecasting and ordering rule creates on its own, and which remedies remove which cause. It builds on the inventory models of the operations core (reorder point, safety stock, order-up-to levels): read mba/mba-inventory-card first, because this card uses its notation and does not re-teach it. It is for MBA students taking a supply chain elective and for managers in retail, distribution, manufacturing and healthcare supply who see erratic orders from customers or send them to suppliers. One example is fully solved, one is half solved and one is left for you, and the self-test answers sit in a separate key. Source basis: Forrester (1961) for the dynamics, Sterman (1989) for the behavioural evidence, Lee, Padmanabhan and Whang (1997) for the four causes, and Chen, Drezner, Ryan and Simchi-Levi (2000) for the formula, all written up in original words with fictional data (Northfield Outdoor Co. and Tallow Creek Medical Center are invented organisations). Replace every [bracketed] value and every row marked Example.

Document control

When to use it

One-sentence rule: use this card whenever the orders a stage receives vary more than the demand of the final customer, and you have to decide whether the cause lies in forecasting, lead times, batching, rationing or pricing.

The idea in plain language

Picture a chain of four stages: a store, the retailer's distribution centre, a wholesaler and a factory. Shoppers buy roughly the same number of headlamps every week. Yet the factory sees orders that jump from very large to almost nothing. That amplification of order variability as orders travel upstream is the bullwhip effect: a small flick at the handle becomes a large swing at the tip.

Forrester (1961) showed with simulation models that delays and feedback in a production-distribution system create such swings. Sterman (1989) ran the beer distribution game with managers and students and found that players under-weight the orders already in the pipeline: they keep ordering when the shelf is low even though shipments are on the way, and the overshoot follows. Lee, Padmanabhan and Whang (1997) then showed that the effect also arises when every manager acts rationally on the information each one has, and named four causes:

1. Demand signal processing. Each stage forecasts from the orders it receives, not from final demand, and resets its order-up-to level every period. A small rise in demand raises the forecast, and the stage orders the demand plus the increase in its target stock. The longer the lead time, the bigger the adjustment.
2. Order batching. Stages order in full trucks, pallets or at month end because each order has a fixed cost (the EOQ logic in mba/mba-inventory-card). Steady demand becomes lumpy orders.
3. Price variations. Promotions and quantity discounts make buyers forward-buy: they buy far more than they need when the price is low and nothing afterwards.
4. The rationing game. When supply is short and the supplier allocates in proportion to orders, customers inflate their orders to get a bigger share, then cancel when the shortage ends. The supplier reads the phantom orders as demand.

Behaviour adds a fifth driver on top of the four: the under-weighting of the supply line that Sterman documented. It is not a separate cause in the 1997 classification, but it is why the effect in a classroom game is larger than any formula predicts.

The cost is real: upstream stages hold more safety stock, build capacity for peaks that never come from final demand, pay overtime and idle time in turn, and deliver worse service because they plan on noise.

Formulas and notation

Measuring the effect. For one stage, compare the variance of the orders it places with the variance of the demand it faces, over the same periods:

$BW > 1$ means the stage amplifies. When the mean of orders and the mean of demand differ over the window (a growing business, a stock build), use the coefficient-of-variation ratio $BW_{CV}$.

The amplification created by forecasting and lead times (Chen, Drezner, Ryan and Simchi-Levi 2000). Demand $D_t$ is independent from period to period. At the end of each period the stage forecasts mean demand as the moving average of the last $p$ observations, sets an order-up-to level equal to $L$ periods of forecast demand, and orders what it sold plus the change in that level:

With a safety-stock term based on an estimated standard deviation, Chen et al. (2000) show this expression is a lower bound, so real amplification is at least this large. For demand that is correlated over time (an AR(1) process with coefficient $\rho$) the bound becomes $1 + \left(\frac{2L}{p} + \frac{2L^2}{p^2}\right)(1 - \rho^p)$.

Several stages (same paper). If every stage $i$ with lead time $L_i$ forecasts from the orders it receives (decentralised), the amplification multiplies. If every stage forecasts from final customer demand (centralised information), the lead times add inside one expression:

Sharing data reduces the effect but does not remove it: lead times still amplify. Both expressions are lower bounds. In the centralised case the bound is exact when no safety-stock term is estimated. The decentralised product is a strict lower bound even then, because the orders an upstream stage receives are correlated over time; the true amplification without shared data is larger, so the value of sharing is larger than the product suggests (Worked example A3).

The cost upstream. The supplier sizes its safety stock on the variability it sees (mba/mba-inventory-card):

Conventions. Chen et al. place an order at the end of period $t$ that arrives $L$ periods later; some textbooks count the review period inside $L$. Check your course convention: it changes the numbers, not the logic.

Excel equivalents. Variance ratio: =VAR.S(orders)/VAR.S(demand). CV ratio: =(STDEV.S(orders)/AVERAGE(orders))/(STDEV.S(demand)/AVERAGE(demand)). Formula: =1+2*L/p+2*L^2/p^2.

Assumptions and limits

The formula is a lower bound under the stated policy. It explains why a longer lead time and a shorter forecasting window amplify, and why sharing demand data helps. It does not capture batching, promotions or rationing, which you measure directly.

How managers use it

Finance cares because amplified orders cost working capital and capacity upstream; the supplier usually prices that cost into its offer. The Cachon and Fisher (2000) analysis of shared information is a useful caution: in their model, faster and cheaper order processing and smaller batches were worth more than the information itself, so test which remedy pays before you sign a data-sharing programme.

Diagnosing the cause

Start with data, then look for the pattern each cause leaves in the order history: steps in orders after demand changes (signal processing), regular spikes and zeros (batching), spikes before price changes (promotions) and spikes during shortages followed by cancellations (rationing).

Worked example A (fully solved): Northfield headlamp chain

Situation (Example, fictional). Northfield Outdoor Co. sells the Trailhead 400 headlamp in its stores and online. Weekly customer demand is independent from week to week, with mean $\mu_D$ = 500 and standard deviation $\sigma_D$ = 80 units. The Northfield distribution centre (DC) forecasts with a 4-week moving average ($p$ = 4) and its supplier delivers in $L$ = 2 weeks. The supplier's own factory plans components with a 4-week moving average of the orders it receives and a 3-week lead time.

A1. One stage.

The DC does nothing wrong: it forecasts sensibly and orders up to two weeks of forecast demand. Yet the supplier sees orders whose variance is 2.5 times that of customer demand.

A2. How the forecasting window and the lead time drive it. Variance ratio $1 + 2L/p + 2L^2/p^2$ for each pair (=1+2*L/p+2*L^2/p^2):

Two lessons. A longer forecasting window smooths the forecast and damps the bullwhip, at the cost of reacting more slowly to real changes in demand. And the ratio depends only on $L/p$: doubling the lead time is undone by doubling the window, which is why cutting lead times is such a powerful remedy.

A3. Two stages: what sharing demand data is worth.

Sharing point-of-sale data cuts the factory's variance ratio to 6.63: about 27% below the decentralised bound of 9.06, and about 40% below the exact 11.03 for this rule (the factory's order is $2.625D_t - 2D_{t-4} + 0.375D_{t-8}$ once the DC rule is substituted into the factory rule). It does not bring it back to 1: the lead times still amplify. The rest of the gain has to come from shorter lead times or longer forecasting windows.

A4. A step in demand. Customer demand rises from 500 to 550 a week in week 7 and stays there. The DC follows the rule above ($p$ = 4, $L$ = 2); the factory follows it on the DC's orders ($p$ = 4, $L$ = 3). Weeks before week 1 are at 500. Orders: DC =D_t+(2/4)*(D_t-D_{t-4}), factory =q_t+(3/4)*(q_t-q_{t-4}).

A 10% rise in demand becomes a 15% jump at the DC and a 26.25% jump at the factory, followed by an undershoot to 531.25 while the factory's forecast catches up. Add batching and a second distributor and the swing becomes a production crisis. This is the pattern teams produce in the beer game.

A5. Measuring batching from data. Weekly demand at one store cluster for camp-fuel canisters and the DC's orders, which go out in full pallets of 1,200:

The moving-average formula would predict a ratio of a few units; the measured ratio is over 300. Almost all of it is batching. The remedy is in logistics (smaller, mixed pallets, more frequent deliveries), not in forecasting.

A6. What it costs the supplier. The supplier holds safety stock for a 95% cycle service level with a 3-week lead time and a holding cost of USD 4 per unit per year.

The amplified orders raise the supplier's safety stock by 132.45 units, a factor of 1.5811 ($\sqrt{2.5}$), on one item. Across thousands of items, and in the capacity the factory keeps for peaks, this is the money that funds a data-sharing or VMI programme.

Worked example B (half solved): Tallow Creek gloves

Situation (Example, fictional). Tallow Creek Medical Center's central store replenishes nitrile gloves to the wards. Ward consumption is independent from week to week with mean 400 boxes and SD 60. The central store forecasts with a 2-week moving average and orders from a distributor with a 1-week lead time. The distributor forecasts from the hospital's orders with a 2-week moving average and a 2-week lead time from the manufacturer.

1. Variance ratio at the central store: $1 + 2(1)/2 + 2(1)^2/2^2$ = [ ] (=1+2*1/2+2*1^2/2^2); SD of its orders = [ ] boxes (=60*SQRT([ ])).
2. The store moves to a 6-week moving average. New ratio = [ ]; new SD = [ ] boxes.
3. Back to $p$ = 2. Distributor stage alone: ratio = [ ]. Decentralised bound at the distributor = [ ] x [ ] = [ ]; SD of distributor orders = [ ] boxes.
4. If the distributor plans on ward consumption data (centralised, total lead time 3 weeks, $p$ = 2): ratio = [ ]; SD = [ ] boxes.
5. Which remedy does more for the distributor here: sharing consumption data, or the longer window in step 2? Explain in two sentences, and name one clinical risk of a longer forecasting window for a critical supply.

Exercise C (unsolved): a supplier's complaint

A regional distributor tells Northfield that its weekly orders for trekking-pole tips are "impossible to plan" (Example, fictional). Northfield's store sales and its orders to the distributor for eight weeks:

1. Compute the mean and SD of demand and of orders, the variance ratio and the CV ratio.
2. Which of the four causes fits this pattern best? What evidence in the table supports it?
3. Propose two remedies, say who in each firm must agree, and how you would re-measure success.
4. Separately, for the forecasting part of Northfield's process ($L$ = 2 weeks), what is the shortest moving-average window $p$ that keeps the variance ratio from forecasting at or below 1.5? What does Northfield give up by choosing it?
5. The distributor proposes vendor-managed inventory. List three data items and two contract terms you would need before agreeing.

Connections to other cards

• mba/mba-inventory-card (core): safety stock, reorder points and the order-up-to logic this card builds on; EOQ is the source of batching.
• mba/mba-operations-guide (core): the session on supply chain coordination and the beer game.
• mba/mba-supply-chain-guide: where this card sits in the elective; contracts, risk pooling and the Northfield mini-case.
• mba/mba-network-design-card: the number and place of stocking points decides lead times, and lead times drive the amplification.
• mba/mba-sourcing-procurement-card: contract terms (minimum order quantities, discounts, allocation rules) that create or remove batching and rationing.
• mba/mba-supply-chain-resilience-card: shortages trigger the rationing game; resilience planning should include allocation rules.
• mba/mba-pricing-card and mba/mba-revenue-management-card: promotions and price changes that cause forward buying.
• academic/stats-time-series-arima: forecasting methods and autocorrelation ($\rho$ in the AR(1) form).

Common mistakes and misconceptions

• Comparing orders with orders. The bullwhip ratio compares the orders a stage places with the demand it faces. A supplier comparing its own orders with its customer's orders measures only its own stage.
• Blaming irrational planners. Lee, Padmanabhan and Whang (1997) show the effect arises with rational behaviour; fix the structure (information, lead times, batching, pricing, allocation), not only the people.
• Expecting data sharing to remove the effect. Centralised information still leaves the lead-time term (Worked example A3).
• Using the formula for batching. The variance formula covers forecasting with lead times only; batching and promotions must be measured from data (Worked example A5).
• Mixing sample and population variance. VAR.S and VAR.P differ; use the same one for orders and demand.
• Ignoring seasonality. A seasonal pattern in demand inflates both variances; compare deseasonalised series or like-for-like periods.
• Using the variance ratio when means differ. In a growing business or a stock build, compare coefficients of variation instead.
• Treating a longer forecasting window as free. It damps the bullwhip but reacts late to real changes, a cost in fashion and in critical supplies.
• Counting on-hand stock instead of inventory position. Orders in the pipeline must be counted, or planners re-order what is already on the way (Sterman 1989).
• Assuming the effect is always bad to the same degree. Some smoothing costs service; the goal is lower total cost across the chain, which needs a contract that shares the gains.

Self-test

1. Define the bullwhip effect in one sentence and name the four causes from Lee, Padmanabhan and Whang (1997).
2. A stage forecasts with a 4-period moving average and has a 1-period lead time. Compute the variance ratio.
3. Same window, lead time 4 periods. Compute the ratio and explain the change.
4. Two stages each with $L$ = 1 and $p$ = 4. Compute the ratio at the second stage with decentralised and with centralised information.
5. Demand SD is 100 units a week. With the ratio of question 2, what is the SD of the stage's orders?
6. Why does the ratio depend only on $L/p$?
7. A stage's orders have SD 90 and its demand SD 30, with equal means. What is the variance ratio?
8. Which cause does each remedy address? (a) Allocation on past sales. (b) Everyday low pricing. (c) Mixed-product trucks. (d) Sharing point-of-sale data.
9. What behavioural finding did Sterman (1989) report from the beer distribution game?
10. When should you use the CV ratio instead of the variance ratio?
11. Cachon and Fisher (2000) compared the value of shared information with other improvements. What did they find?
12. Interleaving: which card first? (a) A hospital asks how many vaccine doses to buy before flu season. (b) A supplier sees orders that spike every fourth week. (c) A retailer asks how many DCs it needs.

My notes and next steps

Sources and further reading

• Forrester, J. W. (1961). Industrial Dynamics. MIT Press.
• Sterman, J. D. (1989). Modeling managerial behavior: misperceptions of feedback in a dynamic decision making experiment. Management Science, 35(3), 321-339.
• Lee, H. L., Padmanabhan, V., and Whang, S. (1997). Information distortion in a supply chain: the bullwhip effect. Management Science, 43(4), 546-558.
• Lee, H. L., Padmanabhan, V., and Whang, S. (1997). The bullwhip effect in supply chains. Sloan Management Review, Spring 1997. Reprinted in IEEE Engineering Management Review (2015), 43(2), 108-117.
• Chen, F., Drezner, Z., Ryan, J. K., and Simchi-Levi, D. (2000). Quantifying the bullwhip effect in a simple supply chain: the impact of forecasting, lead times, and information. Management Science, 46(3), 436-443.
• Chen, F., Ryan, J. K., and Simchi-Levi, D. (2000). The impact of exponential smoothing forecasts on the bullwhip effect. Naval Research Logistics, 47(4), 269-286.
• Cachon, G. P., and Fisher, M. (2000). Supply chain inventory management and the value of shared information. Management Science, 46(8), 1032-1048.
• Dejonckheere, J., Disney, S. M., Lambrecht, M. R., and Towill, D. R. (2003). Measuring and avoiding the bullwhip effect: a control theoretic approach. European Journal of Operational Research, 147(3), 567-590.
• Textbooks (further reading, by name only): Simchi-Levi, Kaminsky and Simchi-Levi, Designing and Managing the Supply Chain; Chopra and Meindl, Supply Chain Management: Strategy, Planning, and Operation; Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World.
• Open resources: MIT OpenCourseWare, 15.762J Supply Chain Planning (Spring 2011), https://ocw.mit.edu/courses/15-762j-supply-chain-planning-spring-2011/, and 15.871 Introduction to System Dynamics (Fall 2013), https://ocw.mit.edu/courses/15-871-introduction-to-system-dynamics-fall-2013/.

Answer key

Worked example B. (1) Ratio 2.5000; SD 94.87 boxes (=60*SQRT(2.5)). (2) $p$ = 6: ratio 1.3889 (=1+2*1/6+2*1^2/6^2); SD 70.71 boxes. (3) Distributor alone ($L$ = 2, $p$ = 2): ratio 5.0000. Decentralised bound: 2.5 x 5 = 12.5000; SD 212.13 boxes (=60*SQRT(12.5)). (4) Centralised, $L$ = 3, $p$ = 2: ratio 8.5000 (=1+2*3/2+2*3^2/2^2); SD 174.93 boxes. (5) Sharing consumption data cuts the distributor's ratio from 12.5 to 8.5; the longer window at the store alone cuts the store's ratio from 2.5 to 1.39 and, decentralised, the distributor's from 12.5 to 6.94 (1.3889 x 5). Both help; together with a longer window at the distributor they help most. Clinical risk: a long window reacts slowly to a real surge (an outbreak), so pair it with an override rule for critical items.

Exercise C. (1) Demand: mean 304.38, SD 17.41, variance 303.125. Orders: mean 300.00, SD 277.75, variance 77,142.86. Variance ratio 254.49 (=VAR.S(orders)/VAR.S(demand)); CV ratio 16.19. (2) Order batching: orders come only in multiples of 300 with zero weeks in a regular three-week cycle, while demand barely moves. (3) Example remedies: smaller order multiples or mixed cartons (logistics and the distributor's warehouse must agree); a fixed weekly delivery slot with electronic ordering (procurement and the distributor's sales team). Re-measure the variance ratio and the distributor's forecast error over the next 8 to 12 weeks. (4) $1 + 4/p + 8/p^2 \le 1.5$ gives $p \ge 4 + \sqrt{32}$ = 9.66, so $p$ = 10 weeks (ratio 1.4800). Northfield reacts slowly to real changes in demand, a risk for seasonal items. (5) Data: daily or weekly store sales by item, on-hand and in-transit stock, promotion calendar. Terms: service level and stock ownership (who owns stock at the store until sale), limits and penalties for overstock or stock-outs.

Self-test.

1. The growth of order variability as orders move upstream. Causes: demand signal processing, order batching, price variations, rationing game.
2. $1 + 2/4 + 2/16$ = 1.6250.
3. $1 + 8/4 + 32/16$ = 5.0000. The order-up-to level covers four periods, so every change in the forecast is multiplied by four in the order.
4. Decentralised: 1.625 x 1.625 = 2.6406 (a lower bound; the exact value for this rule is 2.8359). Centralised: total lead time 2, so 2.5000.
5. $100\sqrt{1.625}$ = 127.48 units.
6. The order adds $(L/p)(D_t - D_{t-p})$ to demand: lead time scales the adjustment and the window divides it.
7. $90^2 / 30^2$ = 9.00.
8. (a) Rationing game. (b) Price variations. (c) Order batching. (d) Demand signal processing.
9. Players under-weighted the supply line (orders placed but not yet received), so they over-ordered and created large oscillations even though final demand changed only once.
10. When the mean of orders and the mean of demand differ over the measured window, as in growth or stock building.
11. In their model, shared demand information lowered costs, but faster and cheaper order processing (shorter lead times, smaller batches) lowered them more.
12. (a) Newsvendor (mba/mba-inventory-card). (b) This card, batching. (c) mba/mba-network-design-card.

Revision history

Standard and licence note

Based on the published supply chain literature, chiefly Forrester (1961), Sterman (1989), Lee, Padmanabhan and Whang (1997) and Chen, Drezner, Ryan and Simchi-Levi (2000); all text, data and examples are original and fictional, and textbooks are named as further reading only. The beer distribution game is named as a teaching exercise; no game material is reproduced. This template is not affiliated with or endorsed by any business school, publisher or course. Check your course materials for the lead-time convention your professor uses. Template licence: MIT.