Stocking Retail Assortments Under Dynamic Consumer Substitution

Transcription

1 Stocking Retail Assortments Under Dynamic Consumer Substitution Siddharth Mahajan Garret van Ryzin Operations Research, May-June 2001 Presented by Felipe Caro Seminar: Theory of OM April 15th, 2004 This summary presentation is based on: Mahajan, Siddharth, and Garrett van Ryzin. "Stocking Retail Assortments Under Dynamic Consumer Substitution." Operations Research 49, no. 3 (2001).

2 Motivation Sample path with T sequential customers Retailer RETAILER Customer t with preferences U t =(U t0, U t1,, U tn ) Category with n substitutable variants

3 Motivation Retail consumers might substitute if their initial choice is out of stock Retailer s inventory decisions should account for substitution effect Consumers' final choice depends on what he/she sees available on the shelf. In most previous models demand is independent of inventory levels. Contribution of this paper: Determination of initial inventory levels (singleperiod) taking into account dynamic substitution effects

4 Outline Brief Literature Review Model Formulation Structural Properties Component-wise Sales Function Total Profit Function Continuous Model Optimizing Assortment Inventories Numerical Experiments Price and Scale Effects Conclusions

5 Brief Literature Review Demand models with static substitution: Smith and Agrawal (2000) / van Ryzin and Mahajan (1999): 1. First choice is independent of stock levels 2. If first choice is out of stock, the sale is lost (no second choice) 3. Consequently, demand is independent of inventory levels Papers that model dynamic effects of stockouts on consumer behavior Anupindi et al. (1997): concerned solely with estimation problems (no inventory decisions) Noonan (1995): only one substitution attempt

6 Model Formulation Notation: (See "Model Formulation" on page 336 of the Mahajan and van Ryzin paper)

7 Model Description Assumptions: Sequence of customers is finite w.p.1 Each customer makes a unique choice w.p.1 Some special cases: -Multinomial Logit (MNL): -Markovian Second Choice: = = j [1] j q P( Ut Ut ) -Universal Backup: all customers have an identical second choice -Lancaster demand: attribute space [0,1] and customer t has a random ideal point L t, then U = u + ξ j j j t t P( U = U U = U ) = p U > U > > U [2] k [1] j k 0 [3] [ n] t t t t j t t t U = a b L l j t t j

8 Model Formulation Profit function: h j (x,w) = sales of variant j given x and w. Individual profit: p j (x,w) = p j h j (x,w) c j x j. Total profit: p(x,w) = Sp j (x,w). Retailer s objective: max E[ π ( x, ω)] x 0 Recursive formulation: System function: Sales-to-go: ( t, t) f( x, ) d x U t Ut = xt e = x t + 1 η ( x, ω) = x f ( x, U ) + η ( x, ω) j j j j t t t t t t+ 1 t+ 1 Border conditions: η j T+ 1( T+ 1, ω) = 0 1 = x x x

9 Structural Properties Lemma 1: x y ï x t y t for all sample paths. Decreasing Differences: h:sµqö satisfies decreasing differences in (z,q) if h(z,q) h(z,q) h(z,q ) h(z,q ) for all z z, q q. Lemma: if max{h(z,q): zœs} has at least one solution for every q œq, then the largest maximizer z * (q) is nonincreasing in q.

10 Structural Properties Theorem 1: The function h j (z e j + q,w) satisfies: (a) Concavity in z for all w. (b) Decreasing differences in (z,q) for all w. Corollary 1: (a) A base-stock level is optimal for maximizing the component-wise profits. (b) The component-wise optimal base-stock level for j is nonincreasing in x i (i j). ï Usual newsboy problem

11 Structural Properties Let: T(x,w)= h j (x,w) = total sales H j (z,w)= T(z e j + q,w) If all variants have identical price and cost, then: p(x,w) = p T(x,w) c x j Theorem 2: There exists initial inventory levels x and sample paths w for which: (a) T(x,w) is not component-wise concave in x. (b) H j (z,w) does not satisfy decreasing differences in (z,q).

12 Structural Properties Counterexample for (a): (See the first two tables on page 339 of the Mahajan and van Ryzin paper) Continuous model: Customer t requires a quantity Q t of fluid with distribution F t ( ) Redefine sample paths: w = {(U t, Q t ): t=1,,t} Theorem 3: There exist sample paths on which p(x,w) is not quasi-concave

13 Optimizing Assortment Inventories Lemma 3: If the purchase quantities Q t are bounded continuous random variables then E[h(x,w)] = E[ h(x,w)] Calculating η(x,ω): (See the equations and explanation in section 4.1, page 341 of the Mahajan and van Ryzin paper)

14 Optimizing Assortment Inventories Sample Path Gradient Algorithm (See the steps in section 4.2, page 341 of the Mahajan and van Ryzin paper) Theorem 4: If F t ( ) is Lipschitz for all t then any limit of y k is a stationary point

15 Numerical Experiments Heuristic policies (with T~Poisson) Let q j (S) be the probability of choosing variant j from S 1. Independent Newsboy: demand for each variant is independent of stock on hand. x j = λq j ( S) + z j 2 λq j ( S) j S I 2. Pooled Newsboy: customers freely substitute among all available variants. j qs ( ) = q( S) xs ( ) = λqs ( ) + z 2 λqs ( ) j S j j q ( S) xp = x( S) qs ( )

16 Numerical Experiments Assumptions: -T~Poisson(30). -Q t ~exp(1). Example 1: Assume MNL utilities: j j i q ( S) = P( U = max{ U : i S}) = j where υ = e t t i u / µ j e j S u / µ 0 j = 0 i S j υ 0 υ + υ Equal cost and prices. Result from van Ryzin and Mahajan: assume v 1 v 2 v, the n optimal assortment is a set A k ={1,2,,k} with k=1,,n. Simulation: profit within ±1% with 95% confidence. Several starting points tested.

17 Numerical Experiments Performance Comparison for Example 1. -Independent newsboy is biased: underestimates popular items, over estimates less popular items. (See Table 2 and Figure 1 on pages of the Mahajan and van Ryzin paper.)

18 Numerical Experiments Example 2: Two variants: variant 1 less popular but with high margin, and variant 2 more popular but lower margin. Sample Path Gradient policy induces customers to upgrade to the highmargin variant. (See Table 3 and Figure 4 on page 345 of the Mahajan and van Ryzin paper.)

19 Numerical Experiments Example 3: Lancaster demand model: U = a b L l t t j L t ~U[0,1], a=0.2, b=1 (See Table 4, Figures 5, and Figure 6 on pages of the Mahajan and van Ryzin paper.)

20 Price and Scale Effects Measure of evenness : y is more fashionable than z, if z majorizes y: Observations: n n [] i [] i i= 1 i= 1 k y = k [] i [] = 1,, 1 i= 1 i= 1 z y i z k n 1) If v is more fashionable than w, then w is more profitable 2) Price or volume increase ï higher variety is offered

21 Conclusions Concluding remarks General choice model. Improvement upon the existing literature. Interesting numerical experiments with valuable insights. Comments Assortment or inventory problem? Assumes full information. Single replenishment: price decision might be relevant. Industry evidence (field study).