A Parsimonious Model of Stock-keeping-Unit (SKU) Choice Teck H. Ho Haas School of Business UC, Berkeley Joint work with Juin-Kuan Chong, NUS The Goal Search for best-fitting model in SKU choice January 2004 Teck H. Ho Berkeley 1 Purchase History of Panelist 14110874 Purchase SKU Occasion Bought Brand COKE PEPSI ROYAL Size TAB CROWN PRIVATE 9.00 7.50 4.50 3.75 3.00 2.25 1.06 LABEL 1 73 0 1* 0 0 0 0 0 1* 0 0 0 0 2 73 0 2* 0 0 0 0 0 2* 0 0 0 0 3 73 0 3* 0 0 0 0 0 3* 0 0 0 0 4 69 1* 3 0 0 0 0 0 4* 0 0 0 0 5 102 2* 3 0 0 0 0 0 4 0 1* 0 0 6 99 2 4* 0 0 0 0 0 4 0 2* 0 0 7 73 2 5* 0 0 0 0 0 5* 0 2 0 0 8 73 2 6* 0 0 0 0 0 6* 0 2 0 0 9 99 3* 6 0 0 0 0 0 6 0 3* 0 0 10 99 4* 6 0 0 0 0 0 6 0 4* 0 0 11 14 4 7* 0 0 0 1* 0 6 0 4 0 0 12 73 4 8* 0 0 0 1 0 7* 0 4 0 0 13 3 5* 8 0 0 0 2* 0 7 0 4 0 0 14 73 5 9* 0 0 0 2 0 8* 0 4 0 0 15 61 6 9 0 0 0 3* 0 8 0 4 0 0 CLASSIC DIET CAFFEINE CAFFEINE FR CRISP FREE CLASSIC Purchase SKU Occasion Bought Flavor 1 73 0 0 1* 0 0 2 73 0 0 2* 0 0 3 73 0 0 3* 0 0 4 69 0 0 3 1* 0 5 102 0 0 3 2* 0 6 99 0 0 4* 2 0 7 73 0 0 5* 2 0 8 73 0 0 6* 2 0 9 99 0 0 6 3* 0 10 99 0 0 6 4* 0 11 14 0 1* 6 4 0 12 73 0 1 7* 4 0 13 3 1* 1 7 4 0 14 73 1 1 8* 4 0 15 61 1 1 8 5* January 2004 0 Teck H. Ho Berkeley 2 The consumer choice setting Household i chooses a product (or stock-keeping-unit (SKU)) from a choice menu on a series of purchase incidences indexed by t Before each purchase incidence, each SKU j is characterized by a set of marketing-mix activities: price (P), display (D), and feature advertisement (AD) The modeler also observes: Household i’s SKU choices on purchase incidences 1, 2, …, t-1 January 2004 Teck H. Ho Berkeley 3 Research question To develop a good descriptive model of SKU choice to predict the probability of household i choosing SKU j on purchase incidence t Pij (t ) January 2004 Teck H. Ho Berkeley 4 Criteria of a “good” model - Specification Simple (i.e, small number of parameters) Model complexity does not increase with number of items/feature levels in the choice menu Increasing number of items Satisfies plausible principles of human behavior Incorporate psychological findings into model Fits and predicts choice behaviors well (e.g., Guadagni and Little, 1983; Fader and Hardie, 1996) January 2004 Teck H. Ho Berkeley 5 Complex menus & increasing number of items Category Bacon Number of SKUs Year 1 Year 2 New Number of Brands Year 1 Year 2 New Number of Sizes Year 1 Year 2 New Number of Flavors Year 1 Year 2 New 49 56 13 21 26 5 7 7 0 11 10 1 Cola 118 134 23 16 15 1 14 13 2 6 8 2 Eggs 35 35 3 10 10 2 3 3 0 5 5 0 Frozen Pizza 298 290 39 32 40 8 138 135 7 63 65 8 Hotdogs 112 121 16 32 37 6 9 11 2 15 15 0 IceCream 376 384 45 35 37 2 8 9 1 131 135 14 Potato Chips 207 235 78 24 26 5 30 32 4 28 29 4 Regular Cereal 197 234 45 29 34 6 63 70 10 43 44 2 Spaghetti Sauce 166 174 28 37 37 4 26 30 4 29 28 2 Yogurt 241 254 47 14 12 1 7 7 0 64 67 10 87 87 19 21 19 0 9 11 2 8 8 0 250 264 71 38 40 3 48 54 14 17 20 4 84 82 24 25 23 2 7 9 2 14 17 3 Soap 207 206 36 46 44 1 39 41 3 16 17 2 Toothpaste 219 237 40 22 24 2 40 40 4 44 50 7 176 186 35 27 28 3 30 31 4 33 35 4 Bathroom Tissue Detergent Paper Towel AVERAGE January 2004 Teck H. Ho Berkeley 6 Criteria of a “good” model – Estimation Does not aggregate choice (i.e., at the SKU level) Heterogeneity across products (biased estimates); avoid “average” variables; inventory planning Does not throw away observations Choice-based sampling (biased estimates) (Ben-Akiva and Lerman, 1985) January 2004 Teck H. Ho Berkeley 7 Criteria “violations” Model specification Model complexity Many models have complexity increases with number of items Plausible principles of behavior Few attempts to incorporate findings from psychological research in consumer behavior Model estimation Aggregate choice Violation examples: Brand-size combination; “other” product Throwing away observations Violation examples: Top n SKUs; ignore SKUs that have few purchases January 2004 Teck H. Ho Berkeley 8 Notations Examples Household i SKU j Purchase Occasion t Attribute k Attribute level l Panelist 14110874 UPC 11200000847 July 17, 97 Brand COKE For estimation, every product category is assumed to have three attributes (Brand, Size, Flavor) January 2004 Teck H. Ho Berkeley 9 Utility specification U ij (t ) Vij (t ) M ij (t ) ij (t ) Utility = intrinsic value + value associated with marketing-mix activities Error structure captures serial correlations in attribute-level and product-specific utilities Uses latent class to capture heterogeneity No product or attribute-level specific intercept terms! January 2004 Teck H. Ho Berkeley 10 Intrinsic Value K Lk Vij (t ) Aikl (t ) I jkl Aij (t ) k 1 l 1 Intrinsic value consists of both product-specific and attribute-level experiences Example: SKU 14 = {PEPSI, 9.0, DIET}, Panelist = Grace VGrace,14 (t ) AGrace, PEPSI (t ) AGrace,9.0 (t ) AGrace, DIET (t ) AGrace,14(t ) January 2004 Teck H. Ho Berkeley 11 Marketing-mix response M j (t ) P Pj (t ) jD D j (t ) jAD ADij (t ) Control for price, display, and feature advertisement on local newspapers Display and feature advertisement are dummy variables Actual price paid (incorporating coupons and etc.) January 2004 Teck H. Ho Berkeley 12 An overview of the model Productspecific Experience Previous Cumulative Reinforcement Incremental Reinforcement Intrinsic Value Size Utility Attributelevel Experience Marketing-mix Response Consumption Shopping Previous Cumulative Reinforcement Consumption Brand Flavor Incremental Reinforcement Shopping •Intrinsic value consists of both product-specific and attribute-level experiences • Consumption and shopping experiences depend on product and attribute-level familiarity January 2004 Teck H. Ho Berkeley 13 Cumulative attribute-level reinforcement, Aikl(t) Aikl (t ) k Aikl (t 1) Rikl (t ) Cikl (t 1) Sikl (t ), if (k,l) was chosen in t - 1 Rikl (t ) Sikl (t ), otherwise Cumulative attribute-level reinforcement = Decayed previous reinforcement + immediate incremental reinforcement Incremental reinforcement consists of consumption as well as “shopping” experience for chosen level and “shopping” experience only for unchosen levels January 2004 Teck H. Ho Berkeley 14 An overview of the model Productspecific Experience Previous Cumulative Reinforcement Incremental Reinforcement Intrinsic Value Size Utility Attributelevel Experience Marketing-mix Response Consumption Brand Aikl (t ) Flavor Shopping Previous Cumulative Reinforcement Aikl (t 1) Consumption Cikl (t 1) Incremental Reinforcement Rikl (t ) Shopping Sikl (t ) January 2004 Teck H. Ho Berkeley 15 Consumption (Cikl(t-1)) & shopping (Sikl(t)) experiences Cikl (t 1) Ck 0 Ck 1 Fikl (t 2) S ikl (t ) S k 0 S k 1 Fikl (t 1) Consumption & shopping experiences depend on consumer’s familiarity with the level “Shopping” experience because people care about foregone utilities from actions/products that they could have chosen (Camerer and Ho, 1999) Ck1 < 0 captures “law of diminishing marginal utility” Sk1 > 0 captures “memory-based decision making” (Alba, Hucthinson, and Lynch, 1991) January 2004 Teck H. Ho Berkeley 16 Variety-seeking behavior Cikl (t ) Ck 0 Ck 1 Fikl (t 1) Sikl (t ) S k 0 S k 1 Fikl (t ) Modeled as negative reinforcement (e.g., Lattin, 1987) Under our model setup, it is driven by Ck1 < 0 (satiation) (Erdem, 1992; McAlister, 1982) or Sk1 > 0 (“grass is greener” effect) (Kahn, 1998) January 2004 Teck H. Ho Berkeley 17 Product and attribute-level familiarities Fikl (t ) ln( 1 a Tikl (t )) Fij (t ) ln ( 1 θ p Tij(t)) Product and attribute-level familiarity is concave in number of times the product and attribute levels are consumed (Tikl(t) & Tij(t)) Also tried linear and step functions Log function fits best and is also the most appealing conceptually January 2004 Teck H. Ho Berkeley 18 Main ideas Utility consists of intrinsic value and value associated with marketing-mix response Intrinsic value has two components: product-specific and attribute level experiences Incremental reinforcement has both consumption and shopping experience, which depends on product and attribute-level familiarity Each unchosen attribute level receives a different “shopping” reinforcement The model has 5 6 ( K 1) parameters for a Kattribute product category Example: K=3 (brand, size, flavor), the model has 29 parameters January 2004 Teck H. Ho Berkeley 19 Data Set Panel-level market basket data set 124,000 product purchases across 15 product categories (10 food + 5 non-food) Purchases made by 513 households at 5 stores located within the same neighborhood over a 2-year period + Data from Fader and Hardie (1996) January 2004 Teck H. Ho Berkeley 20 Data Set Category Summary Total Sample Size Number of Households Number of SKUs Total Number of Levels in All Salient Attributes 1 Number of Parameters Our Model Fader and Hardie Guadagni and Little Flavor/Ingredient Total Number Example January 2004 Large Product Categories Egg 9903 9781 14590 14705 3698 12218 4111 7022 12594 4226 5214 2993 7171 12978 482 38 594 59 528 106 429 141 314 62 495 108 334 128 382 285 356 288 320 194 384 243 306 259 471 321 480 242 20 22 40 41 45 53 64 95 96 102 107 119 124 153 59 75 91 73 79 133 59 115 227 59 117 297 59 125 139 59 141 231 59 163 271 59 225 585 59 227 591 59 239 403 59 249 501 59 273 533 59 283 657 59 341 499 10 Downy Snuggle Bounce 21 Scottissue Northern Charmin 17 Coca Cola Pepsi Ryl Crown 26 Oscar Mayer W.CornKing Lazy Maple 29 Jays Lays Ruffles 15 Dannon Yoplait Kemps 41 Ragu Prego Hunts 47 Dial Dove Ivory 24 Crest Colgate Arm&Hammer 41 Tide Wisk All 35 Kellogg Gnrl Mills Post 4 Small Medium Large 11 4 rl. 1 rl. 12 rl. 16 67.6 oz. 288 oz. 144 oz. 7 16 oz. 12 oz. 24 oz. 34 6.5 oz. 7 oz. 6 oz. 7 6 oz. 8 oz. 32 oz. 30 30 oz. 26 oz. 14 oz. 42 15 oz. 14 oz. 9.5 oz. 44 6.4 oz. 4.6 oz. 6 oz. 62 64 oz. 128 oz. 42 oz. 73 12 oz. 18 oz. 15 oz. Salient Attribute Description Brand Total Number 12 Example Crystal Fm. Prv. Label W.R.Valley Package Size Total Number Example Small Product Categories Bathroom Tissue Cola Bacon Fabric Softener 3 12 ct. 18 ct. 6 ct. 5 Large Extra Large Jumbo Paper Towel Hotdog Potato Chip 27 38 Bounty Oscar Mayer Scottowels Hygrade Versatile W.CornKing 9 1 rl. 3 rl. 2 rl. 11 16 oz. 12 oz. 40 oz. Yogurt Spaghetti Sauce Soap Toothpaste Detergent Regular Cereal 4 8 8 12 17 15 32 74 31 18 Regular Unscented Regular Regular White Paper Beef Regular Plain Plain Regular Staingard Regular Diet Smoked Print Chckn&Pork BBQ Strawberry Itln. Garden Original Light Soft Scented Caffn. Free Hkry Smoked As.Colors Pork&Turkey SC & Onion Raspberry Tmt. & Herb Unscented Teck H. Ho 51 Tartar Ctrl Bk. Soda Regular 21 45 Reg. Liquid Corn Con. Pwd Wheat Bran Reg. Pwd Rice Berkeley 21 Estimation Max LL I ij (t ) Pij (t ) k ; Ck 0 , Ck1 ; S k1 ; p ; C p 0 , C p1 ; S p1 i j t A , p ; X , X P, D, AD Maximize the likelihood of observing the data The first 13 weeks of data for initialization; the next 65 weeks for calibration and the last 26 weeks for model validation Benchmark against Fader and Hardie (1996)’s model January 2004 Teck H. Ho Berkeley 22 FH Model K Lk Vij (t ) [vkl Aikl (t )] I jkl k 1 l 1 Has attribute-level specific terms Does not capture familiarity-based consumption as well as shopping experience S k 1 Ck 1 0 January 2004 Teck H. Ho Berkeley 23 Key Results (Small Categories) Number of parameters Our model = 59 (two-segment models); FH model = 75-163 Comparison was made on small product categories (less than 200 parameters) Calibration The hit probability is 7% better than F&H model Better in every single product category Validation The hit probability is 8% better than F&H model Better in every single product category January 2004 Teck H. Ho Berkeley 24 Key Results (Small Categories) - Calibration Table 2: Calibration and Validation Results for the Small Product Categories Egg Fabric Bathroom Softener Tissue Cola Bacon Paper Towel Hotdog Calibration Sample Size 6252 4417 9303 9241 2383 7768 2577 Log-likelihood Our Model Fader and Hardie Guadagni and Little Empirical Frequency -5414 -5699 -5978 -8691 -2600 -3074 -3039 -15504 -11287 -13196 -30384 -10592 -11911 -34861 -3272 -3523 -3892 -6000 -8845 -9407 -25104 -3635 -3927 -8502 Average Hit Probability Our Model Fader and Hardie Guadagni and Little Empirical Frequency 0.55 0.53 0.53 0.33 0.83 0.82 0.81 0.03 0.51 0.45 0.06 0.60 0.55 0.03 0.37 0.32 0.27 0.12 0.56 0.52 0.05 0.47 0.44 0.06 Adjusted 2 Our Model Fader and Hardie Guadagni and Little 0.37 0.34 0.30 0.83 0.80 0.80 0.63 0.56 - 0.69 0.65 - 0.44 0.39 0.33 0.65 0.62 - 0.57 0.52 - January 2004 Teck H. Ho Berkeley 25 Key Results (Small Categories) - Validation Validation Sample Size 2494 2137 3510 3495 842 2889 927 Log-likelihood Our Model Fader and Hardie Guadagni and Little Empirical Frequency -2262 -2486 -2518 -3781 -1484 -1814 -1650 -7867 -4357 -5346 -12108 -3910 -4527 -12463 -1201 -1283 -1521 -2461 -3194 -3467 -11781 -1445 -1556 -3089 Average Hit Probability Our Model Fader and Hardie Guadagni and Little Empirical Frequency 0.56 0.53 0.55 0.30 0.81 0.80 0.79 0.03 0.50 0.42 0.05 0.58 0.53 0.03 0.39 0.33 0.26 0.11 0.57 0.54 0.04 0.46 0.42 0.06 Adjusted 2 Our Model Fader and Hardie Guadagni and Little 0.39 0.32 0.31 0.80 0.76 0.77 0.64 0.55 - 0.68 0.63 - 0.49 0.43 0.33 0.72 0.69 - 0.51 0.44 - Note 1: FH's fabric softener has 4 attributes, hence the number of parameters are adjusted accordingly. In addition, they initialized with 52 weeks of data which we also adhere to here. January 2004 Teck H. Ho Berkeley 26 Key Results (Large Categories) Table 3: Calibration and Validation Results for the Large Product Categories Potato Chip Yogurt Spaghetti Sauce Soap Regular Cereal Ice Cream Frozen Pizza 4395 7949 2701 3197 1892 4596 8262 4351 3396 -5485 -17868 -6930 -36341 -3076 -10601 -3605 -13007 -1762 -7636 -4287 -20008 -8998 -36301 -2854 -19318 -2496 -14915 Average Hit Probability Our Model Empirical Frequency 0.57 0.03 0.74 0.01 0.60 0.03 0.65 0.02 0.70 0.02 0.71 0.02 0.68 0.02 0.77 0.02 0.76 0.01 Adjusted 2 Our Model 0.69 0.81 0.70 0.72 0.76 0.78 0.75 0.85 0.83 Sample Size 1698 3189 1085 1268 792 1677 3040 1623 1412 Log-likelihood Our Model Empirical Frequency -1309 -7210 -2755 -15202 -1147 -4304 -1269 -5598 -693 -4067 -1036 -10015 -2686 -13265 -880 -7960 -894 -7017 Average Hit Probability Our Model Empirical Frequency 0.65 0.04 0.73 0.01 0.60 0.03 0.67 0.02 0.72 0.02 0.79 0.02 0.71 0.02 0.80 0.02 0.77 0.01 Adjusted 2 Our Model 0.81 0.81 0.72 0.76 0.82 0.89 0.79 0.88 0.86 Toothpaste Detergent Calibration Sample Size Log-likelihood Our Model Empirical Frequency Validation January 2004 Teck H. Ho Berkeley 27 Tests of Key Behavioral Premises Table 4: Tests of Behavioral Premises: Log-likelihood Ratios of Nested Models Egg Small Product Categories Fabric Bathroom Softener Tissue Cola Bacon Large Product Categories Paper Towel Hotdog Potato Chip Spaghetti Yogurt Sauce Soap Regular Frozen Toothpaste Detergent Cereal Ice Cream Pizza The Full Model (Log-likelihood) -5414 -2600 -11287 -10592 -3272 -8845 -3635 -5485 -6930 -3076 -3605 -1762 -4287 -8998 -2854 -2496 Behavioral Premise 1 No Attribute-level Reinforcement 246* 722* 2370* 936* 568* 1281* 405* 3014* 414* 2110* 1479* 1932* 695* 581* 728* 2598* No Product-level Reinforcement 1080* 176* 3868* 3385* 568* 3699* 1013* 4523* 694* 3400* 846* 2196* 618* 2215* 293* 2748* Behavioral Premise 2 No Consumption Experience 306* 70* 737* 326* 237* 1487* 370* 3226* 1299* 1646* 234* 2202* 309* 230* 44* 533* No Shopping Experience 775* 1027* 2566* 1725* 556* 1751* 449* 5134* 724* 3288* 1801* 2168* 2093* 8038* 1680* 1621* Behavioral Premise 3 No Familiarity 765* 549* 3636* 1498* 1623* 3075* 719* 1763* 696* 2259* 3469* 2074* 955* 4813* 293* 645* Note 1: * indicates significance at 1%. January 2004 Teck H. Ho Berkeley 28 Conclusion Our model Simple but fits and predicts better Neither aggregates choice nor discards data Shows both product and attribute-level experiences matter Shows consumers accumulate both shopping and consumption experiences IRI has implemented this model at a leading consumer packaged goods firm January 2004 Teck H. Ho Berkeley 29