Demand Forecasting with Stockouts and Substitution at The Cornell Store Vishal Gaur, Joonkyum Lee, Suresh Muthulingam Acknowledgment: Amr Farahat, Gary Swisher, Deb Barth, Mike Staurowsky for presentation at the Revenue Management Conference, Georgia Tech, May 21-22, 2012 Demand Forecasting with Stockouts and Substitution at The Cornell Store Vishal Gaur, Joonkyum Lee, Suresh Muthulingam Acknowledgment: Amr Farahat, Gary Swisher, Deb Barth, Mike Staurowsky for presentation at the Revenue Management Conference, Georgia Tech, May 21-22, 2012 Business Context • Non‐profit organization, member of National Association of College Stores (NACS) • Textbooks – New/Used/Rental, Custom/Standard – Sourced from publishers, wholesalers, students • Competition – Online stores (Amazon.com), peer‐to‐peer market in used books, local stores around campus • Scale of business – 3,500 books per semester for about 2000 course‐instructor combinations – 2 buyers Gaur, Lee, Muthulingam / Cornell University 3 Research Objectives Estimate substitution behavior between new and used textbooks Improve demand forecast and stocking decisions • Unique setting – Only censored demand (sales) data are available. – Occurrence of stockouts is known, but the time of occurrence of stockout is not known. – The total (potential) number of customers is known (enrollment number). – Show impact from a pilot implementation at the Cornell Store Gaur, Lee, Muthulingam / Cornell University 4 Literature Review Estimation of demand with stockout‐based substitution • • • • • • Anupindi, Dada and Gupta (1998) Kӧk and Fisher (2007) Fisher and Vaidyanathan (2009) Musalem et al. (2010) Conlon and Mortimer (2010) Vulcano, van Ryzin and Ratliff (2012) Stockout decisions with stockout‐based substitution • Smith and Agrawal (2000) • Mahajan and van Ryzin (2001) • Honhon, Gaur and Seshadri (2010) Gaur, Lee, Muthulingam / Cornell University 5 Data Description ISBN Semester Estimated Enrollment New Sales Used New Used Sales Stockout Stockout New Price # of Courses 0‐00‐648595‐2 54 300 82 88 0 0‐00‐648595‐2 54 300 55 135 0‐00‐945000‐B 54 160 97 0‐02‐019985‐6 54 150 0‐02‐019985‐6 54 0‐02‐076200‐3 Avg books % per course Required 0 18.95 2 11 1 1 18.95 3 44 1 0 25.99 14 82 0 1 220 36 59 0 55 12 3 5 0‐02‐332200‐4 55 54 8 0‐02‐351320‐9 55 300 0‐02‐352960‐1 55 0‐02‐360760‐2 Level Department 0 1 art_science 14.66667 1 1 hotel 2 3 1 2 management 12 1 19 1 3 art_science 0 12 1 14 1 2 art_science 0 1 8 1 11 1 2 art_science 12 0 0 52.5 1 7 1 4 architecture 92 108 1 1 10 3 18 1 1 art_science 30 11 10 0 0 10.75 1 7 1 3 ecology 55 80 22 34 0 0 6 2 7 1 2 architecture 0‐02‐402150‐4 56 60 14 34 0 0 9.5 1 8 1 3 ilr 0‐02‐427691‐X 56 80 24 26 1 1 104 1 1 1 4 management 0‐02‐427691‐X 56 65 11 19 0 0 107.25 1 1 1 3 art_science 0‐02‐428810‐1 56 600 124 198 0 0 46 1 10 1 2 ilr • Sales are observed; demand is not. • There is substitution between new and used books when one of them stocks out. Gaur, Lee, Muthulingam / Cornell University 6 Data Description continued Number of ISBNs in each semester Semester Spring 05 Fall 05 Spring 06 Fall 06 Spring 07 Fall 07 Spring 08 Fall 08 Spring 09 Fall 09 Required books 3229 3283 3281 3395 3353 3261 3112 3135 2874 3080 Optional books 488 475 586 468 460 432 415 406 353 412 Number of ISBNs per course‐instructor‐semester Mean Std Dev Median Max 99% 95% Min Required books 3.21 2.93 2 27 13 9 1 Optional books 2.04 1.98 1 29 10 6 1 Gaur, Lee, Muthulingam / Cornell University 7 Forecasting Model • Customer’s utility from {new, used}, Unj = xj + random noise – x: forecasting variables or product attributes • price, course level, department, required/optional, # of other books in course – β: coefficients or weights of product attributes (parameters to be estimated) • Choice probability is given by a multinomial logit formulation pj S Gaur, Lee, Muthulingam / Cornell University exp x j 1 exp xnew exp xused 8 Timeline illustrating the sequence of events. One value of total sales can be obtained in many ways. Time of stock‐out not observed Used book stocks out New book stocks out NEW NEW NEW USED USED USED New and used books are available ‐ Some students buy new books ‐ Some students buy used books ‐ Some students buy nothing New book sales Used book sales Gaur, Lee, Muthulingam / Cornell University Total number of customers Time line Only new books are available No books ‐ Some students buy new books ‐ Some students substitute new books for used books ‐ Some students buy nothing ‐ No sales Total new book sales Total used book sales New book sales 9 Enumerate sample paths to account for stockouts and substitution (new/used books) • Maximize likelihood max • i observations Pr textbook i' s sales observed sales Probability that we observe a given sales of a textbook – Example: Both new and used books stock out. – Pr(New book sales=Sn, Used book sales=Su | Number of enrollment=N) Case 1: used books stock out first and new book stocks out Case 2: new books stock out first and used book stocks out Sn 1 N Sn Su k 0 j 0 Su 1 N Sn Su k 0 j 0 Su 1 k j ! k S u j P P P Su 1 ! k! j! n u 0 Sn 1 k j ! S n k j P P P Sn 1 ! k! j! n u 0 Sn new book sales k used book sales j students buy nothing Assortment: new and used book Beginning of sales Gaur, Lee, Muthulingam / Cornell University N Sn Su j l 0 N Sn Su j l 0 Sn 1 k l ! P Sn 1 k ! l! n,u Sn k Su 1 k l ! P Su 1 k ! l! u,n P0,u l Su k P0,n l Su‐k used book sales l students buy nothing Assortment: used books only New book stock‐out Used book stock‐out No books N customers 10 Evaluation of our method against alternative demand forecasting methods • Alternative Method 1. Use only uncensored observations – Use only those observations in which no stockout occurred) – Analyze low demand observations only. • Alternative Method 2. Ignore stockout and substitution – Assume demand = sales. – Example: Both new and used books stock out. Pr(New book sales=Sn, Used book sales=Su | Number of enrollment=N) = Pn S n Pu S u P0 N Sn Su • Alternative Method 3. Uncensoring without modeling substitution – Account for stockout information. – After a stockout occurrence, do not account for substitute demand. Gaur, Lee, Muthulingam / Cornell University 11 A simulation experiment to evaluate the demand forecasting methods Given demand model (KNOWN ) % GAP Compare results with true demand model Generate random observations • Product characteristics • Inventories • Customer Arrivals • Sales occurrence Estimate parameters using our model and alternative models • Estimates of mean demand • Estimates of choice parameters Gaur, Lee, Muthulingam / Cornell University 12 Simulation Setup Six parameters (β) to be estimated • Constant, Price, 1000Level for new and used books • True value of β = (‐1.2, ‐0.2, 0.7, ‐1, ‐0.1, 0.4) Random values for product attributes (x) • Enrollment number ~ 500*beta(2,18) • Normalized Price ~ Normal(1,0.2) • 1000Level ~ Bernoulli(0.3) • 10,000 observations are generated Stocking levels • Stocking level with respect to the expected demand={0.5, 0.75, 1, 1.25, 1.5, 1.75, 2} • We try these 7 different stocking levels to assess the effect of stockouts on forecast accuracy and computation time. Gaur, Lee, Muthulingam / Cornell University 13 MPE in expected demand (New and used books) Result 1: Evaluation of demand forecast with respect to the true mean 30% 20% 10% 0% ‐10% ‐20% ‐30% ‐40% ‐50% ‐60% ‐70% ‐80% ‐90% ‐100% Mean Percentage Error in Expected Demand 0.5 (96.6%) 0.75 (89.3%) 1 (61.6%) 1.25 (28.0%) 1.5 (11.5%) 1.75 (5.3%) 2 (2.3%) Stocking Level as a fraction of expected demand (overall stockout rate shown in parentheses) Model 1: Use only uncensored observations Model 2: Ignore stockout and substitution Model 3: Ignore substitution Our model: account for stockout and substitution With stockout information, true demand can be estimated even when there is a high incidence of stockouts. Gaur, Lee, Muthulingam / Cornell University 14 Result 2 (Parameter Recovery): Comparison of estimated parameters with the true values % gap between true parameters and estimates obtained by different methods for stocking level = 0.75 (stockout rate = 98.3%) New Books Used Books Price elasticity Whether 1000 level ‐1.0 ‐0.1 0.4 0.72 2.6% ‐1.01 ‐1.6% ‐0.08 17.3% 0.40 0.3% ‐0.24 ‐18.2% 0.81 15.4% ‐0.89 11.3% ‐0.11 ‐13.0% 0.52 29.2% ‐1.68 ‐40.0% ‐0.18 7.6% 0.64 ‐8.9% ‐1.47 ‐47.4% ‐0.08 19.8% 0.33 ‐18.4% ‐2.0 ‐66.5% ‐0.43 ‐116.1% 0.67 ‐4.1% ‐1.77 ‐77.3% ‐0.20 ‐100.9% 0.32 ‐20.1% Intercept Price elasticity True Parameters ‐1.2 ‐0.2 0.7 Account for stockout and substitution ‐1.17 2.4% ‐0.22 ‐9.8% Ignore substitution ‐1.04 13.6% Ignore stockout and substitution Use only uncensored observations Gaur, Lee, Muthulingam / Cornell University Whether Intercept 1000 level 15 Price elasticity of demand has a larger estimation error when stockouts are more frequent True values: ‐0.2 for new books and ‐0.1 for used books. New book price elasticity Stocking level 0.5 (stockout rate) (96.6%) Account for stockout and substitution ‐0.14 Ignore substitution ‐0.17 0.75 1 1.25 1.5 (89.3%) (61.6%) (28.0%) (11.5%) 1.75 (5.3%) 2 (2.3%) ‐0.22 ‐0.21 ‐0.23 ‐0.21 ‐0.20 ‐0.23 ‐0.24 ‐0.23 ‐0.23 ‐0.22 ‐0.20 ‐0.24 Ignore stockout and substitution ‐0.17 ‐0.18 ‐0.22 ‐0.23 ‐0.21 ‐0.20 ‐0.24 Use only uncensored observations 6.21 ‐0.43 ‐0.24 ‐0.23 ‐0.21 ‐0.20 ‐0.24 1.75 (5.3%) 2 (2.3%) Used book price elasticity Stocking level 0.5 (stockout rate) (96.6%) Account for stockout and substitution ‐0.13 Ignore substitution ‐0.15 0.75 1 1.25 1.5 (89.3%) (61.6%) (28.0%) (11.5%) ‐0.08 ‐0.14 ‐0.13 ‐0.11 ‐0.08 ‐0.10 ‐0.11 ‐0.15 ‐0.13 ‐0.12 ‐0.08 ‐0.09 Ignore stockout and substitution ‐0.07 ‐0.08 ‐0.12 ‐0.13 ‐0.11 ‐0.08 ‐0.09 Use only uncensored observations 11.75 ‐0.20 ‐0.09 ‐0.12 ‐0.10 ‐0.08 ‐0.09 Gaur, Lee, Muthulingam / Cornell University 16 Pilot implementation • 3 groups to compare performance – Group 1: business as usual (control group) – Group 2: we provide demand forecasts – Group 3: we provide suggested stocking level • 30 sets of matching textbooks for each group – Drop 6 sets because of dropped courses or changes in parameters. – Use 24 sets (72 textbooks), accounting for 3.4% of the books for the Spring 2012 semester. Gaur, Lee, Muthulingam / Cornell University 17 Matching process proportion of the sets with identical attributes average (median) maximum difference Price of a new book (V) 8.3% 7.1% (6.8%) Number of courses (Q) 95.8% 2.8% (0.0%) Average number of books per course (B) 95.8% 1.3% (0.0%) Proportion required (R) 100.0% 0.0% (0.0%) Enrollment (E) 91.7% 2.1% (0.0%) New ISBN (W) 91.7% ‐ Course level (L) 100.0% ‐ Odd semester (O) 100.0% ‐ Department (D) 100.0% ‐ Gaur, Lee, Muthulingam / Cornell University 18 Demand information provided for group 2 1. If new and used books are stocked in plenty Expected Demand Standard Deviation 2. If only new books are stocked in plenty 3. If only used books are stocked in plenty Expected Demand Standard Deviation Expected Demand Standard Deviation Group Set ISBN New book Used book Total New book Used book New book New book Used book Used book 2 1 9780205711499 1.2 4.3 5.5 1.1 1.8 1.6 1.2 4.6 1.8 2 2 9780030327162 5.9 28.2 34.1 2.3 4.3 9.2 2.8 30.4 4.3 2 3 9781556520747 1.8 7.2 9.0 1.3 2.1 3.0 1.6 8.0 2.1 2 4 9780141441474 14.7 61.5 76.3 3.6 6.0 25.0 4.6 68.2 6.1 2 5 9780195042399 1.4 7.7 9.1 1.1 2.1 2.5 1.5 8.4 2.1 2 6 9780679721888 1.5 7.0 8.5 1.2 2.1 2.5 1.5 7.6 2.1 2 7 9781566564151 2.1 7.8 9.9 1.4 2.1 3.7 1.7 8.8 2.1 2 8 9780071546058 5.0 16.0 21.0 2.1 3.3 7.5 2.5 17.8 3.3 2 9 9780679723417 4.7 15.2 19.9 2.1 3.2 7.1 2.4 16.9 3.2 2 10 9780822200161 9.9 33.0 42.9 3.0 4.7 15.1 3.6 36.8 4.8 Gaur, Lee, Muthulingam / Cornell University 19 Process of finding suggested stocking level for group 3 Step Example Provide initial suggested stocking level For set 1 30 new books 70 used books base on the ABS heuristic and neighborhood search using simulation Update availability information Usually used‐book availability is limited Other attributes (e.g., price or enrollment) might change Provide revised suggested stocking level Solve the optimization problem with restrictions on availability (and others) Gaur, Lee, Muthulingam / Cornell University Only 50 used books are available 40 new books 50 used books 20 Result of the pilot experiment: Comparison of realized profit Total Profit Total profit Group 1 Group 2 Group 3 1611.59 1747.03 1775.48 Profit Difference Group 2 ‐ Group 1 Group 3 ‐ Group 1 Group 3 ‐ Group 2 Mean paired difference in profit 5.6431 6.8285 1.1854 p‐value of paired t‐test 0.0711 0.0938 0.4280 Median paired difference in profit 7.2137 3.5463 ‐0.2625 p‐value of Wilcoxon signed‐rank test 0.0678 0.0476 0.4527 p‐value of sign test 0.0758 0.0320 0.5000 Group 2 and 1 Group 3 and 1 Group 3 and 2 17.80% (10.41%) 6.33% (‐1.71%) * The p values are one‐tailed Percentage Gap Mean (median) percentage gap in profit 11.89% (16.68%) Both group 2 and group 3 perform significantly better than group 1 Gaur, Lee, Muthulingam / Cornell University 21 Result of the pilot experiment 2: Sales to stock ratio and stockout rate Sales to stocking level ratio Mean (median) sales to stocking level ratio Total New books Used books Group 1 Group 2 Group 3 0.60 (0.59) 0.42 (0.40) 0.75 (0.88) 0.58 (0.61) 0.55 (0.50) 0.70 (0.75) 0.72 (0.85) 0.86 (1.00) 0.70 (0.80) • Total and new‐book sales to stock ratio of group 3 is significantly higher than those of other groups • Total sales to stock ratios of group 1 and group 2 are similar but they have different configuration Stockout rate Total New books Used books Group 1 12.50% 7.14% 17.86% Group 2 12.50% 7.14% 17.86% Group 3 39.29% 50.00% 28.57% • Stockout rate of group 3 is significantly higher than those of other groups Gaur, Lee, Muthulingam / Cornell University 22 Result of the pilot experiment 3: Drivers – Stocking level and product mix Difference in stocking level and product mix Group 1 Group 2 Group 3 Mean (median) total stocking level to estimated enrollment ratio 0.53 (0.50) 0.56 (0.53) 0.38 (0.41) Mean (median) used‐book stocking level to total stocking level ratio 0.46 (0.51) 0.55 (0.72) 0.70 (0.82) • The total and the relative used‐book stocking level of group 1 is substantially lower than those of other groups • The relative used‐book stocking level of group 2 is considerably lower than that of group 3 Gaur, Lee, Muthulingam / Cornell University 23 Reasoning – Difference between group 1 and 2 • Different product mix: – Group 1 stocks relatively more new books and group 2 stocks relatively more used books. – When none of new and used books stocks out: overage cost of a new book is high profitability of group 1 is worse – When either type of book stocks out substitution starts to occur substitution rate from used books to new books is higher than that from new books to used books (even though margin of a new book is better) price cost salvage overage cost underage cost critical ratio new book 1 0.6 0.48 0.12 0.4 0.769 used book 0.75 0.375 0.3 0.075 0.375 0.833 Note. The numbers are normalized to set new‐book price to 1. The numbers are averaged ones over all book titles. Gaur, Lee, Muthulingam / Cornell University 24 Reasoning – Difference between group 1 and 3 • Total stocking level of group 1 is high – Stocking level from solving newsvendor problems for new and used books independently is too high when critical fractile is high (Honhon et al. 2010) • Impact of ignoring the substitution rate – Usually the availability of used books is limited. – The buyers tend to make up for the shortfall of used books by buying additional new books to match the initial total stocking level. – However, the substitution rate from used books to new books is typically low (0.2 ~ 0.4) Gaur, Lee, Muthulingam / Cornell University 25 Summary • Accurate demand forecasts can be obtained by incorporating stockout information in the model even if the stock‐out rate is significantly high(~90%). • Enrollment (traffic) data are valuable for estimation. • Pilot implementation shows the interdependence between stocking levels of the new and used books. • Splitting the dataset improves computational efficiency, but increases error. ‐ Use safety stock to compensate for the error. • Ongoing work ‐ Extending the analysis to more than two products Gaur, Lee, Muthulingam / Cornell University 26 Gaur, Lee, Muthulingam / Cornell University 27 Summary • Our method recovers parameters more accurately than other methods. – Even if the stock‐out rate is significantly high(~90%), accurate demand forecasts can be obtained by incorporating stockout information in the model. • Pilot implementation shows that our method can make significant improvement in profitability. – Accounting for substitution effect drives different stocking level and product mix. • Our method can be applied dynamically under the restrictions on product availability. • Splitting the dataset improves computational efficiency, but increases error. – Use safety stock to compensate for the error. • Enrollment (traffic) data are valuable for estimation. • Further work – Extending the model to more than two products Gaur, Lee, Muthulingam / Cornell University 28 Data Description - 2 • 61,032 observations across 10 semesters – – – – – – Course, Course level, Instructor, ISBN, Required/Optional text New book price, Used book price, New book cost Estimated and actual enrollment To Provide (Stock‐up‐to level) Sales of new and used books, Returns of new and used books Publisher, Vendor, Buyer code • Contextual information – Types of contracts with different publishers and wholesalers Gaur, Lee, Muthulingam / Cornell University 29 Customer Purchase Process Buy new NEW USED p q Buy used 1-p-q Don’t buy any When new books are out‐of‐stock: NEW q' > q USED Buy used 1-q Don’t buy any An analogous process occurs when used books are out‐of‐stock. Gaur, Lee, Muthulingam / Cornell University 30 Challenges in estimating the forecasting model • Demand is not observed, only sales are. This is due to stockouts. • Number of books (of the other type) sold before a stockout occurs is not known. Only total sales are known. Gaur, Lee, Muthulingam / Cornell University 31 Identifying the opportunity for improvement by estimating lost sales and excess inventory • • The demand forecast model and the store’s ordering decisions can be used to estimate excess inventory and lost sales Picture shows estimation results for 100 test observations (model estimated on 2,361 observations). 0.9 0.8 Estimated excess inventory and lost sales as % of total enrollment Excess Inventory/Enrollment Lost Sales/Enrollment 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 ‐0.1 0 Gaur, Lee, Muthulingam / Cornell University 0.2 0.4 0.6 0.8 1 1.2 1.4 Total inventory of new and used books as % of total enrollment 32 Stock-out rates also differ between new and used textbooks, showing substitution patterns Didn’t order used book Didn’t order new book N/A Stocked out Ordered new Didn’t stock book out Ordered used book N/A Stocked out Didn’t stock out 0% 1% 5% 6% 7% 8% 8% 22% 34% 21% 16% 71% 41% 30% 29% 100% • Substitution: Most but not all students prefer to buy used books. In 8% of cases, the bookstore stocked out of new books, but not old. • Profit margin on new books = 25% Profit margin on used books = 35%. • Used books stock out about half the time. New books stock out about 25% of the time. Thus, the bookstore could benefit by stocking more used books, if available. Gaur, Lee, Muthulingam / Cornell University 33 0.8 Excess Inventory/Enrollment Lost Sales/Enrollment 0.6 0.4 0.2 0 ‐0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 ‐0.4 ‐0.6 ‐0.8 Gaur, Lee, Muthulingam / Cornell University 34 Objectives of the study For the store For us Assess the current performance of the bookstore and find areas for improvement Devise a method for demand forecasting and improve ordering of new and used books Add to algorithm in the system Improve competitiveness in sourcing used books from students Improve effectiveness of a brick & mortar store competing with online channels Advancement in forecasting and planning Issues related to managerial behavior in operational decisions Modeling prices and supply in the sourcing of used books Performance effects of the contracts between the store and publishers/wholesalers Gaur, Lee, Muthulingam / Cornell University 35 Example: One Text for Multiple Courses DEPT COURSE SECTION TITLE AUTHOR INSTRUCTOR REQ/OPT EST ENROLL. QTY ACT ENROLL. TOTAL SALES GOVT 1101 103 ELEMENTS OF STYLE STRUNK PEPINSKY R 17 202 18 145 GOVT 1101 102 ELEMENTS OF STYLE STRUNK KEMERLI R 17 202 18 145 GOVT 1101 104 ELEMENTS OF STYLE STRUNK BENSEL O 17 202 17 145 PHIL 1111 102 ELEMENTS OF STYLE STRUNK KYLE O 17 202 18 145 PHIL 1112 104 ELEMENTS OF STYLE STRUNK STAPLETON R 17 202 18 145 ENGL 1127 102 ELEMENTS OF STYLE STRUNK TORRES R 17 202 17 145 ENGL 1158 109 ELEMENTS OF STYLE STRUNK VAUGHN O 17 202 18 145 ENGL 1185 109 ELEMENTS OF STYLE STRUNK MCQUEEN‐ THOMSON R 17 202 16 145 ILRIC 2301 1 ELEMENTS OF STYLE STRUNK COOK R 17 202 11 145 ENGL 2700 101 ELEMENTS OF STYLE STRUNK SCHWARZ R 17 202 14 145 ENGL 2710 102 ELEMENTS OF STYLE STRUNK FRIED R 18 202 18 145 GOVT 3303 1 ELEMENTS OF STYLE STRUNK TURNER O 200 202 53 145 ILRIC 4330 1 ELEMENTS OF STYLE STRUNK TURNER O 200 202 129 145 ILRIC 6330 1 ELEMENTS OF STYLE STRUNK TURNER O 12 202 14 145 8 R 600 202 379 145 14 courses Gaur, Lee, Muthulingam / Cornell University 36 Example: One Text for Many Courses DEPT COURSE SECTION TITLE AUTHOR INSTRUCTOR REQ/OPT EST ENROLL. QTY ACT ENROLL. TOTAL SALES GOVT 1101 103 ELEMENTS OF STYLE STRUNK PEPINSKY R 17 202 18 145 GOVT 1101 102 ELEMENTS OF STYLE STRUNK KEMERLI R 17 202 18 145 GOVT 1101 104 ELEMENTS OF STYLE STRUNK BENSEL O 17 202 17 145 PHIL 1111 102 ELEMENTS OF STYLE STRUNK KYLE O 17 202 18 145 PHIL 1112 104 ELEMENTS OF STYLE STRUNK STAPLETON R 17 202 18 145 ENGL 1127 102 ELEMENTS OF STYLE STRUNK TORRES R 17 202 17 145 ENGL 1158 109 ELEMENTS OF STYLE STRUNK VAUGHN O 17 202 18 145 ENGL 1185 109 ELEMENTS OF STYLE STRUNK MCQUEEN‐ THOMSON R 17 202 16 145 ILRIC 2301 1 ELEMENTS OF STYLE STRUNK COOK R 17 202 11 145 ENGL 2700 101 ELEMENTS OF STYLE STRUNK SCHWARZ R 17 202 14 145 ENGL 2710 102 ELEMENTS OF STYLE STRUNK FRIED R 18 202 18 145 GOVT 3303 1 ELEMENTS OF STYLE STRUNK TURNER O 200 202 53 145 ILRIC 4330 1 ELEMENTS OF STYLE STRUNK TURNER O 200 202 129 145 ILRIC 6330 1 ELEMENTS OF STYLE STRUNK TURNER O 12 202 14 145 8 R 600 202 379 145 14 courses Gaur, Lee, Muthulingam / Cornell University 37 Example contd.: Time-series data for “Elements of Style” Sales Semester Estimated Enrollment Qty Actual Enrollment 54 281 217 228 95 104 199 62 221 188 174 93 43 136 64 407 292 339 111 93 204 72 256 183 173 84 40 124 74 312 193 278 101 90 191 82 356 223 271 75 77 152 84 290 233 259 95 91 186 92 252 118 133 28 37 65 94 600 202 379 84 61 145 Gaur, Lee, Muthulingam / Cornell University 38 Another Example: One Course with Multiple Texts Course / Dept 1101 PSYCH Title Author Reqd/ Optional Instructor Estimated Enrollment Qty Actual Enrollment Total Sales MBTI FORM M 6165 ‐ R MAAS 1355 1117 1095 845 POWER SLEEP MAAS R MAAS 1280 938 1020 712 FRONTIERS OF PSYCHOLOGY MAAS R MAAS 1280 1050 1020 825 ICLICKER R MAAS 10775 4985 9164 3662 O MAAS 1280 100 1020 98 R MAAS 1280 1100 1020 701 ICLICKER POWER NAP KIT COLLEGE GELB EDITION 3 PK: PSYCHOLOGY W/ SG, MYERS PSYCH PORTAL Gaur, Lee, Muthulingam / Cornell University 39 Performance Evaluation • % Gap between true expected demand and forecasted expected demand True Draw N, price, 1000Level observations Estimation using MLE ˆ1 , ˆ2 , ˆ3 Draw N, price, 1000Level for estimation True Gaur, Lee, Muthulingam / Cornell University % GAP ˆ1 , ˆ 2 , ˆ 3 40 Numerical Experiment: Simulation Process There are 7 stocking levels. For each stocking level, we have 10 sets of observations. A set of observations consists of 1,000 textbooks. Draw random enrollment, price, and 1000level Identify stock‐out information and sales of new and used books Gaur, Lee, Muthulingam / Cornell University Compute choice probabilities, expected demand, and inventory level Find estimates of parameters Each customer observes current assortment and makes purchasing decision Adjust remaining inventory Sample a new set of 100,000 textbooks Find expected demand using true beta and estimates 41 Planning cycle for the Spring semester Mid‐Nov Mid‐Dec Semester begins Mid‐Feb time Selling season Faculty place requests for books for the Spring semester Source used books from students and wholesalers Gaur, Lee, Muthulingam / Cornell University Procure new books Monitor stock outs and get expedited shipments as needed Manage returns 42 Result 2 (Parameter Recovery): Comparison of estimated parameters with the true values % gap between true parameters and estimates obtained by different methods for stocking level = 0.75 (stockout rate = 98.3%) New Books Used Books Price elasticity Whether 1000 level ‐1.0 ‐0.1 0.4 0.72 2.6% ‐1.01 ‐1.6% ‐0.08 17.3% 0.40 0.3% ‐0.24 ‐18.2% 0.81 15.4% ‐0.89 11.3% ‐0.11 ‐13.0% 0.52 29.2% ‐1.68 ‐40.0% ‐0.18 7.6% 0.64 ‐8.9% ‐1.47 ‐47.4% ‐0.08 19.8% 0.33 ‐18.4% ‐2.0 ‐66.5% ‐0.43 ‐116.1% 0.67 ‐4.1% ‐1.77 ‐77.3% ‐0.20 ‐100.9% 0.32 ‐20.1% Intercept Price elasticity True Parameters ‐1.2 ‐0.2 0.7 Account for stockout and substitution ‐1.17 2.4% ‐0.22 ‐9.8% Ignore substitution ‐1.04 13.6% Ignore stockout and substitution Use only uncensored observations Gaur, Lee, Muthulingam / Cornell University Whether Intercept 1000 level 43 Result 3: Taming the computation time # of textbooks Computation Time (sec) # of iterations 100 29 37 500 232 36 1,000 422 33 5,000 2,285 19 10,000 24,251 72 • Computation time increases dramatically with # of observations • Types of observations that are more time‐intensive: – When both new & used books stockout: Likelihood function involves triple summation – When enrollment is large • Splitting the data set improves computational efficiency, but with a small sacrifice of estimation accuracy Gaur, Lee, Muthulingam / Cornell University 44 Performance assessment of the store and differences between buyers Criteria used by buyers to determine stock-up-to levels Stock‐up‐to level for an ISBN is a function of 1. Total estimated enrollment across all courses that request the textbook, 2. Course level, 3. Price of textbook, 4. Whether required/optional, 5. Number of requested books in the course Gaur, Lee, Muthulingam / Cornell University 46 These criteria are confirmed by the data regression results for all ISBNs with enrollment between 100 and 400 (inclusive) Log model Linear model Optional ISBN Required ISBN Optional ISBN Required ISBN # of observations 579 4834 586 4904 R2 51% 61% 21% 54% rmse 0.71 0.45 35.06 42.45 cv of dep var 24.89 10.58 119.33 49.47 1.49*** 1.56*** 18.96*** 41.14*** 0.02* 0.03*** 0.00* 0.00* NewPrice ‐0.36*** ‐0.14*** ‐0.07** ‐0.12*** Actual Enrollment 0.70*** 0.69*** 0.20*** 0.55*** Total # of books per course ‐0.44*** ‐0.11*** ‐1.17*** ‐1.40*** Intercept AvgCourselevel Gaur, Lee, Muthulingam / Cornell University 47 Regression results for the entire data set also confirm the criteria used by the buyers EST1 ToProvide= α + Est_Enrolli*β + Diff_Enrolli*γ + New Pricei* δ + avg booksi* φ + # of coursesi* λ + New ISBNi* η + Controls i* μ + εi ACT1 ToProvide= α + Act_Enrolli*β + New Pricei* δ + avg booksi* φ + # of coursesi* λ + New ISBNi* η + Controls i* μ + εi EST_REQ & ACT_REQ and EST_OPT & ACT_OPT EST1 & ACT1 models for Courses that use Required and Optional Texts Controls ( semester) Coefficient estimates for controls are not provided below! EST1 est_enroll diff_enroll .56*** (.0016) .41*** (.0037) act_enroll new price Avg Books/~e # of courses _Ireqopt_2 New ISBN R_Sq Adj_R_Sq Number ACT1 EST_REQ ACT_REQ .58*** (.0017) .42*** (.0039) .0061 (.0049) .016 (.02) -.55*** (.15) 23*** (.53) 1** (.35) .57*** (.0016) .0036 (.005) -.074*** (.021) 2.3*** (.14) 23*** (.54) .7 (.36) 0.81*** 0.81 35950 0.80*** 0.80 35950 -.0018 (.0051) .0012 (.021) -1.6*** (.15) EST_OPT ACT_OPT .22*** (.003) .25*** (.0056) .59*** (.0017) -.005 (.0053) -.1*** (.022) 1.6*** (.14) 1.6*** (.37) 1.2** (.38) 0.83*** 0.83 31794 0.82*** 0.82 31794 -.054*** (.007) -.22*** (.038) 4.2*** (.24) .22*** (.003) -.053*** (.0071) -.21*** (.038) 3.7*** (.22) 3*** (.54) 3*** (.54) 0.61*** 0.61 4154 0.61*** 0.61 4154 * p<0.05, ** p<0.01, *** p<0.001 Gaur, Lee, Muthulingam / Cornell University 48 Stock-out rates differ between optional and required textbooks A rough benchmark for stockout rate: Margin on a book = 25%, Salvage value = 80% of cost = 60% of price Newsvendor critical fractile = 62.5% Stockout rate = 37.5% Optional books # obs. 5,567 Average Stockout rate 35.3% Required books # obs. 40,898 Average Stockout rate 17.1% Gaur, Lee, Muthulingam / Cornell University 49 There are consistent differences between the two buyers in their stock-up-to levels Qty/ Estimated Enrollment Required Optional Semester Alice Rachel A&R Alice Rachel A&R Alice Rachel A&R 3 0.62 0.56 0.59 0.69 0.63 0.67 0.21 0.19 0.28 4 0.62 0.61 0.64 0.68 0.67 0.69 0.22 0.21 0.23 5 0.60 0.58 0.58 0.65 0.63 0.63 0.21 0.21 0.19 6 0.61 0.60 0.54 0.66 0.65 0.64 0.20 0.19 0.23 7 0.59 0.56 0.57 0.65 0.61 0.60 0.20 0.19 0.21 8 0.62 0.64 0.63 0.66 0.68 0.68 0.21 0.19 0.27 9 0.59 0.59 0.54 0.63 0.63 0.59 0.19 0.21 0.20 10 0.60 0.60 0.64 0.65 0.65 0.67 0.20 0.19 0.12 • Alice stocks more than Rachel in 11 out of 16 cases. • A 1% difference in stocking levels can result in a significant difference in expected profits, esp. when demand uncertainty and in-stock rates are high. Gaur, Lee, Muthulingam / Cornell University 50 Differences between buyers are confirmed in a regression on various control variables EST1 ToProvide/Est Enrollment = α + New Pricei* δ + avg booksi* φ + # of coursesi* λ + New ISBNi* η + Racheli* π + Alice & Racheli* ψ + Controls i* μ + εi ACT1 ToProvide/Act Enrollment = α + New Pricei* δ + avg booksi* φ + # of coursesi* λ + New ISBNi* η + Racheli* π + Alice & Racheli* ψ + Controls i* μ + εi EST_REQ & ACT_REQ EST1 & ACT1 models for Courses which use Required Texts EST_OPT & ACT_OPT EST1 & ACT1 models for Courses which use Optional Texts Controls ( semester) Coefficient estimates for controls are not provided below! EST1 new price -.00035*** (.000042) Avg Books/~e .0053*** (.00018) # of courses -.077*** (.0011) _Ireqopt_2 .4*** (.0049) Rachel -.013*** (.0033) A&R .015** (.0053) New ISBN .051*** (.0033) R_sq R_Sq_A Number 0.36*** 0.36 23187 ACT1 EST_REQ ACT_REQ -.003*** -.0003*** -.0032*** (.00016) (.000046) (.00018) -.0035*** .0059*** -.0037*** (.0007) (.00019) (.00076) -.01* -.081*** -.012* (.0044) (.0012) (.0048) .77*** (.019) .019 -.014*** .015 (.013) (.0037) (.015) -.017 .012* -.026 (.021) (.0058) (.023) -.18*** .05*** -.2*** (.013) (.0036) (.014) 0.10*** 0.10 22672 0.23*** 0.23 20565 0.03*** 0.03 20097 EST_OPT ACT_OPT -.00085*** (.000071) -.003*** (.0004) -.024*** (.0023) -.0015*** (.00025) .00097 (.0014) .0087 (.0079) -.0023 (.0057) .026** (.0092) .037*** (.0056) 0.12*** 0.11 2620 .049* (.02) .044 (.032) -.019 (.02) 0.04*** 0.04 2573 * Comment p<0.05, ** p<0.01, *** p<0.001 1. 2. Rachel stocks less than Alice (Significant in models EST1, EST_REQ, and ACT_OPT) When both buyers are involved stocking levels are higher Gaur, Lee, Muthulingam / Cornell University 51 Further work • Develop a forecasting model using the criteria identified by buyers • Compute stock‐up‐to levels and evaluate performance in the next academic year • Characterize differences between the two buyers Gaur, Lee, Muthulingam / Cornell University 52 Main Takeaways • • • A potential exists for improved pricing policies that adjust for competition and used book supply. The used book market is advantageous to the Store. Alternative buyback models exist that explicitly encourage the purchase of new books from the Store. Next steps: • Test robustness of above insights (alternative demand models, Store expertise, …) • Translate insights into actionable pricing tactics (starting with directional guidelines) Examine the degree of substitutability between new and used books and develop a testable demand function. • Gaur, Lee, Muthulingam / Cornell University 53 Overall Summary • Develop a forecasting model using the criteria identified by buyers • Compute stock‐up‐to levels and evaluate performance in the next academic year • Characterize differences between the two buyers • Test robustness of pricing insights (alternative demand models, Store expertise) • Develop actionable pricing tactics (starting with directional guidelines) • Examine the degree of substitutability between new and used books and develop a testable demand function Gaur, Lee, Muthulingam / Cornell University 54