Borrowing Strength Across Stores: Transfer Learning for Choice-Based Revenue Management

1. Introduction

• Local MNL estimates can be noisy with scarce store data, while pooling reduces variance but can bias estimates when stores have different preferences.The paper proposes shrinkage between local estimation and pooling, with finite-sample analysis of how borrowing should depend on local sample size, feature dimension, and cross-store heterogeneity.Source: §1 · pp. 2–3
• 2.6 percentage points over local estimation and 3.1 percentage points over pooled estimation in average revenue capture; adaptive shrinkage wins in more than 80% of category-store-weeks.The improvements are averaged across category-store-weeks; the improvement over pooling is statistically reliable in every category.Source: §1 · p. 4

2. Literature Review

• The paper extends self-concordant-like analysis to the shared-feature, multi-store MNL setting to derive local quadratic expansions, nonasymptotic concentration bounds, and operational regret bounds.These bounds address nonlinear likelihood loss and nonlinear assortment and pricing decisions.Source: §1 · p. 3, §2 · p. 6

3.1. Store Network and Local Choice Models

• The model has a common product catalog and shared product features across stores, while each store’s customer preferences are represented by its own coefficient vector.The analysis focuses on limited store-level choice data and bounded differences among store parameters.Source: §3.1 · p. 6, §3 · p. 6

3.2. Data Scarcity and Cross-Store Heterogeneity

• Retail data scarcity is especially acute for store-category pairs and short-window or frequently changing categories, while demographics, competition, climate, and format can shift store preferences.Stores are modeled as related but not identical.Source: §3.2 · p. 8

3.3. Assumptions

• The guarantees rely on regularity conditions including independent store samples, balanced store sample sizes, and a minimum local sample size proportional to feature dimension.The stated minimum is n_min ≥ C_0 d.Source: §3.3 · p. 9

4.1. Local Estimation

• Local estimation uses only store-specific data and preserves local heterogeneity, but its variance grows as the local sample shrinks, making downstream decisions unstable.Small changes in estimated choice parameters can change selected assortments or recommended prices.Source: §4 · p. 10, §4.1 · p. 10

4.2. Pooled Estimation

• Full pooling uses the entire chain’s data and is statistically stable, but when stores are heterogeneous its target is systematically biased for each store.It provides a benchmark for the maximum variance reduction from borrowing across the chain.Source: §4.2 · p. 11

4.3. Shrinkage Toward Leave-One-Out Pooled Targets

• The shrinkage estimator moves the local estimate toward a pooled target built from other stores’ data, with its weight spanning local estimation and full pooling.A weight of zero gives the local MLE; a weight of one adopts the leave-one-out pooled estimator.Source: §4.3 · p. 11
• The leave-one-out target excludes store k’s data, keeping it independent of the local estimator so their error contributions can be separated in finite-sample analysis.The full pooled estimator reuses store k’s data, making the two estimators dependent.Source: §4.3 · p. 12

5.1. Local Quadratic Approximation of Population Loss

• The self-concordant-like property of the MNL loss bounds the quadratic-approximation remainder at any sample size, unlike a classical Taylor expansion that requires large samples for negligibility.The expansion separates excess risk into a leading quadratic term and a remainder.Source: §5.1 · pp. 13–14

5.2. The Quadratic-Optimal Weight

• The quadratic-optimal weight projects the gain-to-cost ratio onto [0,1], but can be unreliable at full pooling because deterministic mismatch does not shrink with sample size and can make the remainder offset the quadratic gain.The gain reflects variance reduction, while the cost reflects displacement from the store’s true parameter due to heterogeneity.Source: §5.2 · p. 16

5.3. A Damped Shrinkage Weight

• The damped weight retains a reduction in the leading risk term while limiting systematic bias, sacrificing less relative to the quadratic optimum when the gain-to-cost ratio is small.Its leading term is bounded by a fraction 1/(1 + τ_k) of the optimal value in the interior case, with damping costing more as τ_k grows.Source: §5.3 · p. 17

5.4. Main Finite-Sample Bound

• Shrinkage is most attractive with thin local data and low cross-store heterogeneity; the guarantee applies when global heterogeneity is moderate, and its benefit weakens as heterogeneity grows relative to local sample size.The bound is governed by normalized local sample size ρ_k = n_k/d and global heterogeneity Δ; the results distinguish this guarantee condition from store-specific mismatch, which determines how much an individual store borrows.Source: §5.4 · p. 18
• The damped weight borrows more when a store’s local data are thin or its pooled target is close to its true parameter, while global heterogeneity governs whether the finite-sample guarantee holds.The weight has the form of a ratio of local estimation variance to variance plus squared pooled-target bias.Source: §5.4 · p. 19
• Shrinkage’s guaranteed gain is largest at low heterogeneity and fades as heterogeneity rises or the local sample becomes too thin relative to dimension.Figure 1 shows where Theorem 1 guarantees improvement over local estimation; the store-specific weight adapts to distance from its pooled target, which can be below the uniform heterogeneity bound.Source: §5.4 · p. 19

5.5. Order-Based Plug-In Weights

• When a store has positive structural heterogeneity, its target weight on pooling vanishes as local sample size grows relative to dimension, and the plug-in weight vanishes at the same rate.The persistent heterogeneity bias eventually outweighs the declining local variance.Source: §5.5 · p. 20
• When heterogeneity is zero, the target puts all weight on pooling, but consistency alone may not ensure the plug-in weight reaches it; a sharper estimation-error rate is required.Near zero heterogeneity, that rate is difficult to ensure because the mismatch estimate is a difference between noisy estimators; validation avoids estimating the mismatch directly.Source: §5.5 · pp. 20–21
• The order-based plug-in weight is an explicit implementable family, while cross-validation offers a data-tuned alternative when held-out data are available.The plug-in approach follows from the theory; the validation-based weight is selected directly from held-out likelihood and avoids estimating store heterogeneity.Source: §5.5 · p. 21

6.1. Assortment Decisions

• For assortment decisions, the high-probability regret bound trades local estimation error against pooled error and structural mismatch as the shrinkage weight changes.A larger weight places more emphasis on pooled error and mismatch, and less on local error.Source: §6.1 · p. 22
• In assortment optimization, noisy parameter estimates can change which products are selected, especially when several assortments have similar predicted revenues.The bound includes a mismatch term of order ∆ + ∆^2 because the nonlinear MNL pooled minimizer is not a Euclidean average of store parameters.Source: §6.1 · p. 22

6.2. Pricing Decisions

• For pricing a fixed offered set, the high-probability regret bound has the same local-error, pooled-error, and mismatch components, but reflects errors in valuations and price sensitivity.The propositions use a generic shrinkage weight; a weight minimizing an assortment-regret bound need not minimize a pricing-regret bound.Source: §6.2 · pp. 23–24

7.1. Experimental Design

• The study averages regret over 200 Monte Carlo replications and varies either ∆ or n while holding other parameters fixed; it compares local, pooled, and shrinkage estimators.Source: §7.1 · pp. 24–25, §7 · p. 24

7.2. Varying Cross-Store Heterogeneity

• In Figure 2, pooling outperforms local estimation at low heterogeneity, but its regret rises with ∆ until the curves cross at an intermediate value.Source: Figure 2 · p. 26, §7.2 · p. 26
• The oracle stays near the lower-regret local or pooled curve as heterogeneity changes, while the order-based plug-in follows the same pattern at slightly higher regret.Source: §7.2 · p. 26
• Cross-validation favors heavy borrowing when ∆ is small, then reduces its weight as pooling bias grows, keeping regret near the better extreme.Source: §7.2 · pp. 26–27

7.3. Varying Local Sample Size

• At ∆=1, local-estimator regret falls as n_k grows, while pooled regret stays nearly flat and exceeds local regret in assortment except for the thinnest samples.Source: §7.3 · p. 27, Figure 3 · p. 27
• In Figure 3, the oracle shrinkage benchmark has lower regret than both local and pooled estimators throughout, while plug-in and cross-validated shrinkage stay close.Source: §7.3 · p. 27, Figure 3 · p. 27

8.1. Data and Empirical Setting

• The sample spans several hundred stores and 53 weeks across five women’s apparel categories, with 6.82 million product-store-week observations and 3.54 million units sold.Stocked products with zero sales are retained because they inform demand; product variety and units sold are distinct measures.Source: §8.1 · p. 28, Table 1 · p. 28

8.2. Estimators, Rolling Design, and Decision-Aligned Cross-Validation

• The evaluation uses a rolling 8-week training, 2-week validation, and 2-week test design; cross-validation selects shrinkage by validation revenue capture on a common candidate set.The study measures revenue capture rather than regret because no ground-truth demand model is available; it models conditional product choice within a category.Source: §8 · p. 27, §8.2 · p. 29

8.3. Revenue Capture Results

• Across fifteen category-tercile cells, Shrinkage CV captures 36.84% of realized revenue, versus 34.13% for Store MLE and 33.67% for Pooled MLE, and leads in every cell.The average lifts are 2.71 percentage points over Store MLE and 3.18 percentage points over Pooled MLE.Source: Figure 4 · p. 30, §8.3 · p. 30

8.4. Statistical Reliability of the Gains

• 2.64 percentage points over Store MLE and 3.11 percentage points over Pooled MLE are the overall lifts; Shrinkage CV wins in 86.6% and 83.6% of store-weeks, respectively.Both clustered-bootstrap 95% confidence intervals exclude zero: [2.17,3.10] for Store MLE and [2.55,3.68] for Pooled MLE.Source: Table 3 · p. 32, §8.4 · p. 32
• In the low-volume tercile, the lift over Store MLE averages 0.85 percentage points and its interval includes zero, while the lift over Pooled MLE is 3.13 points with interval [1.60,4.64].The authors note that low-volume store-weeks often rank fewer products, while higher-volume weeks can involve broader assortments and more complex substitution.Source: §8.4 · pp. 32–33

9. Conclusion

• The paper uses MNL shrinkage to selectively borrow information across stores, balancing unreliable local estimates against pooling that can obscure customer-preference differences.Source: §9 · p. 33
• Self-concordant-like properties of the MNL likelihood support finite-sample guarantees for estimation and regret bounds for assortment and pricing decisions.The guarantees characterize how shrinkage depends on local sample size, model dimensionality, and cross-store heterogeneity.Source: §9 · p. 33
• Numerical experiments and an apparel-retailer study demonstrate operational benefits of shrinkage relative to separate estimation and full pooling.Source: §9 · p. 33
• The findings identify information transfer as valuable for revenue management in heterogeneous markets and provide a foundation for studying richer nonlinear choice models and more complex revenue-management settings.Source: §9 · p. 33

EC.1.1. Preliminaries

• For symmetric positive definite matrices, the norm subscript identifies the matrix inducing the norm, with Euclidean, operator, and sub-Gaussian norms as stated exceptions.Source: EC.1.1. Preliminaries · p. 36
• A radius condition on heterogeneity keeps pooled population targets at least ρΘ/2 from the boundary of Θ when ∆≤∆0.The appendix defines ∆0 through RΘ(δ)=ρΘ/2 and states that it depends on primitive quantities.Source: EC.1.1. Preliminaries · p. 36

EC.1.2. Hessian representation and the SCL property

• The MNL losses have deterministic Hessians with a covariance representation and satisfy a 2B self-concordant-like curvature inequality.The appendix uses this property to obtain curvature comparisons along parameter paths.Source: EC.1.2. Hessian representation and the… · pp. 37–40

EC.2. Finite-Sample Concentration of the MLE

• On an interior event, the MLE error is related to the score through a path-average covariance matrix, enabling the concentration argument to control estimator error.Source: EC.2. Finite-Sample Concentration of… · p. 41
• The store-level score at the truth is centered, and its whitened form is isotropic and sub-Gaussian under the stated assumptions.Source: EC.2. Finite-Sample Concentration of… · pp. 41–42
• Under the stated assumptions, the appendix gives finite-sample tail bounds for the store-level MLE and a separate bound for its bias.Source: EC.2. Finite-Sample Concentration of… · pp. 43–44

• The constrained shrinkage minimizer is the Euclidean projection of ζ_k/α_k onto [0,1].The unconstrained minimizer is ζ_k/α_k.Source: EC.3. Auxiliary Results for the… · p. 46
• Under Assumptions 1, 2, 3, and 5 and ∆∈[0,∆0], the Euclidean bound is R∆ = 2∆ + C_h e^(12B∆)∆²; it and the interior-margin conclusion also hold for the full pooled population target.Source: EC.3. Auxiliary Results for the… · pp. 48, 50
• The pooled loss has curvature bounded by c_−(∆)S I_d and c_+(∆)S I_d, where c_−(∆) = me^(−2BR_∆) and c_+(∆) = Me^(2BR_∆).Source: EC.3. Auxiliary Results for the… · p. 50
• The auxiliary analysis establishes a quadratic-improvement bound for shrinkage.Source: EC.3. Auxiliary Results for the… · p. 55
• The bounded parameter region supports the pointwise SCL remainder bound, with the defining remainder quantity finite because ωSCL(a)/a → 1/3 as a ↓ 0.Source: EC.3. Auxiliary Results for the… · p. 57
• For k ∈ Θ, 2B∥z_k∥2 ≤ 2BD_Θ, and ∥z_k∥2 ≤ ∥z_k∥_k/√n_km.Source: EC.3. Auxiliary Results for the… · p. 57
• Taking expectations and using (EC.15) proves the first claim; replacing z_k with e_k and applying the q = 3 case of Lemma EC.11 proves the second.Lemma EC.15(ii) and Lemma EC.11 also bound the last term.Source: EC.3. Auxiliary Results for the… · p. 57

EC.4. Conditions on heterogeneity and C0

• The reference radius ∆0 keeps pooled population targets at a fixed interior margin and makes the ∆-dependent constants uniform on [0,∆0].Source: EC.4. Conditions on heterogeneity and C0 · p. 58
• A sufficiently large normalized sample-size floor C0 makes the required strict inequality hold and ensures the estimator’s interior-event and boundary-probability conditions.Source: EC.4. Conditions on heterogeneity and C0 · p. 58
• Lemma EC.17 establishes that C0 and a positive heterogeneity radius can satisfy the stated requirements simultaneously, with dependence only on the listed primitive quantities.Source: EC.4. Conditions on heterogeneity and C0 · p. 58

EC.5. Proofs of the Main Finite-Sample Results

• Combining remainder bounds and restricting to the certified moderate-heterogeneity region yields the form of the bound stated in the result.Source: EC.5. Proofs of the Main Finite-Sample… · p. 59
• Summing the per-store bounds in Theorem 1 proves the global bound.Source: EC.5. Proofs of the Main Finite-Sample… · p. 59
• The proof establishes a two-sided store-specific order bound.Source: EC.5. Proofs of the Main Finite-Sample… · pp. 60–62
• At the boundary ∆k = 0, consistency of b∆k alone is not sufficient for the stated convergence claim.Source: EC.5. Proofs of the Main Finite-Sample… · p. 62

EC.6. Proofs of the Operational Regret Bounds

• The regret bounds use individual and leave-one-out pooled concentration radii to control estimation error, while pooled structural bias bounds the mismatch term.Source: EC.6. Proofs of the Operational Regret… · pp. 63–64
• Over the bounded price set, revenue for a fixed offered set is Lipschitz in the demand parameter, enabling a price-regret reduction to estimation error.Source: EC.6. Proofs of the Operational Regret… · pp. 63–64

EC.7.1. Synthetic data-generating process

• Synthetic experiments use a multi-store MNL network with a shared catalog, store-specific perturbed coefficients, and an outside option; the baseline has K = 10, N = 15, d = 7, and 50 local observations.Source: EC.7.1. Synthetic data-generating… · p. 65

EC.7.2. Realization-specific oracle shrinkage

• The oracle Shrinkage (w⋆) benchmark forms each store’s estimate before averaging performance across stores and replications.Source: EC.7.2. Realization-specific oracle… · pp. 65–66

EC.7.3. Calibration of the plug-in weight

• The plug-in rule calibrates variance from training data using K times the trace of the pooled-loss Hessian pseudoinverse; its weight is shared across stores per replication, without validation tuning.Source: EC.7.3. Calibration of the plug-in… · p. 66

EC.7.4. Main simulation sweeps

• Two 200-replication sweeps vary heterogeneity and local training size; borrowing matters less as local history grows, while heterogeneity shifts pooling from variance relief toward bias.Source: EC.7.4. Main simulation sweeps · pp. 66–67

EC.7.5. Simulation metrics

• Simulation performance is measured by normalized assortment and pricing regret relative to decisions chosen under the true store-level model.Source: EC.7.5. Simulation metrics · p. 67

EC.7.6. Retail-data implementation

• Retail tests use rolling 8/2/2-week train/validation/test windows and common-support products; validation chooses shrinkage weights, with paired store-category-week lifts bootstrapped over stores.Source: EC.7.6. Retail-data implementation · pp. 67–68

EC.7.7. Additional retail-data robustness checks

• Retail robustness checks repeat revenue-capture tests at 10% and 30% display budgets and inspect validation-selected weights across categories and local-information buckets.Source: EC.7.7. Additional retail-data… · p. 68, Figure EC.1 · p. 68, Figure EC.2 · p. 68

EC.7.8. Sensitivity to Parameter Dimension

• As d rises, pooled assortment regret improves and cross-validated shrinkage leads; pooled pricing regret is U-shaped while local pricing regret worsens sharply.Shrinkage remains competitive, but fixed ∆ makes this a robustness check, not a sharp theoretical comparative-static test, and practical-versus-oracle ranking is calibration-sensitive.Source: EC.7.8. Sensitivity to Parameter… · pp. 69–70, Figure EC.4 · p. 70