Safonov (2024) proposed a two-network neural estimator that maps product characteristics and market variables to item-level demand parameters of a linear demand model, trained by least squares on sales, and showed strong finite-sample performance under the low-price-variation conditions where classical regression-based demand estimation fails. The paper explicitly leaves the estimator's statistical properties open. The problem asks for a rigorous statistical theory of this estimator class: identification, consistency, convergence rates, and a limiting distribution (or a proven failure) for the item-level parameter and elasticity estimates. The source itself records that its evaluation assumes independence of price shocks while real data may exhibit price endogeneity and demand shocks hitting an entire product category simultaneously; what the estimator converges to in those cases belongs to the same question about the estimand. No statistical theory of this architecture is known, so the question remains open.

Background

Demand estimation — recovering the relationship between prices and quantities sold from observational data — is a core problem of empirical economics and retail analytics, because the resulting price elasticities (the percentage change in sales caused by a one-percent price change) drive pricing, assortment, and counterfactual analysis. Classical approaches estimate demand systems for many differentiated products; the most influential is the BLP framework of Berry, Levinsohn, and Pakes (1995), which recovers demand from market shares using instruments to handle price endogeneity (prices being correlated with unobserved demand shocks, as when firms raise prices on products that are unobservably more attractive), and whose modern implementation best practices are surveyed by Conlon and Gortmaker (2020). A spatial semiparametric alternative due to Pinkse, Slade, and Brett (2002) models competition through distances in product-characteristic space. Fisher, Gallino, and Li (2018) validated competition-based dynamic pricing with field experiments, and Bajari, Nekipelov, Ryan, and Yang (2015) compared machine-learning predictors with econometric demand models. A recurring practical obstacle is limited within-product price variation: Safonov (2024) recalls the classical fact that the variance of the estimated price coefficient is inversely proportional to the variance of the residualized price, so when a product's price rarely moves, regression estimates of its elasticity become unreliable.

Safonov (2024) addresses this with a neural estimator consisting of two jointly trained networks. An embedding network FembF_{emb} maps categorical product descriptors (the item's position in the retailer's hierarchical product tree, truncated at level 4 for tractability, plus a unique product identifier) to a 64-dimensional embedding ZiZ_i. A parameter network FFCF_{FC} then maps ZiZ_i together with numerical environmental variables XiX_i (aggregated sales histories, competitors' prices, seasonality indicators) to the parameters θi=(αi,βi)\theta_i=(\alpha_i,\beta_i) of a linear demand function Dθi(Pi)=αi+βiPiD_{\theta_i}(P_i)=\alpha_i+\beta_i P_i. Training minimizes the mean squared error between predicted and observed sales; the price enters only through the loss, never as a network input, which is meant to make the model generalize across price levels. In simulations with logit demand and in 1.5 years of daily data from a large e-commerce retailer (baby food category), the method outperformed a spatial regression benchmark under low price variation, reducing the mean squared error of the initial price-parameter estimates by nearly threefold and producing negative elasticity estimates in 100% of cases against 80% for the regression benchmark. Architecturally related work exists in discrete-choice modeling: TasteNet-MNL (Han, Pereira, Ben-Akiva, and Zegras, 2020) lets a network learn taste parameters inside a multinomial logit model (the workhorse discrete-choice model of consumer demand, in which each alternative's choice probability takes the logistic form), and RUMnet (Aouad and Désir, 2022) represents random-utility choice models (models in which a consumer picks the alternative with the highest utility plus random noise) by networks with approximation and generalization bounds — but neither establishes asymptotic theory for estimated demand parameters.

There is precedent for statistical theory about neural estimators: Chen and White (1999) proved convergence rates and asymptotic normality for nonparametric neural-network regression, and Farrell, Liang, and Misra (2021) established convergence rates for deep ReLU networks in nonparametric regression together with asymptotic normality and valid inference for linear functionals, which Safonov cites as the closest framework. Chen (2007) surveys the general theory of sieve estimation, in which an infinite-dimensional unknown function is approximated by a sequence of finite-dimensional parameter spaces (here, neural networks of growing capacity). None of these covers the present construction, however: it is not a network regressing outcomes on covariates, but a network that outputs the parameters of a parametric demand model which are then fitted through a nonlinear least-squares loss — a sieve M-estimator whose behavior also depends on the near-flat price design it is designed for. The paper's conclusion states as its first future direction 'investigating statistical properties of the model's estimates (convergence and asymptotic normality)', and the text acknowledges that the empirical evaluation assumes independence of price shocks, while real data may exhibit price endogeneity and demand shocks hitting an entire product category simultaneously — issues the classical literature handles with instruments, randomized experiments, or equilibrium models such as BLP.

Problem Statement

For the two-network demand-parameter estimator of Safonov (2024) — parameters θi=(αi,βi)\theta_i=(\alpha_i,\beta_i) predicted as FFC(Femb(Si),Xi)F_{FC}(F_{emb}(S_i),X_i) from product descriptors SiS_i and environmental variables XiX_i, inside the linear demand model Dθi(Pi)=αi+βiPiD_{\theta_i}(P_i)=\alpha_i+\beta_i P_i, trained by least squares on observed sales with the price entering only through the loss — establish a rigorous statistical theory.

Statistical setting. The data are a panel of NN items observed over TT periods (days in the source's application), with observed sales qitq_{it}, price pitp_{it}, categorical product descriptors SiS_i, and environmental variables XiX_i; training minimizes the empirical mean squared error 1n∑i,t(Dθ^i(pit)−qit)2\frac{1}{n}\sum_{i,t}(D_{\hat{\theta}_i}(p_{it})-q_{it})^2, with n=NTn=NT, and the price enters only through this loss. The item-level estimand is the population least-squares minimizer θi∗=(αi∗,βi∗)\theta_i^*=(\alpha_i^*,\beta_i^*) of E[(Dθi(pit)−qit)2]\mathbb{E}[(D_{\theta_i}(p_{it})-q_{it})^2] under item ii's joint distribution of prices and demand; when the affine specification Dθi(Pi)=αi+βiPiD_{\theta_i}(P_i)=\alpha_i+\beta_i P_i does not hold exactly — as in the source's own simulation, where sales arise from a multinomial logit and expected sales are not affine in price — θi∗\theta_i^* is the corresponding pseudo-true (projection) value. The observables used by the theory (raw daily item-level sales, the source's filtered subsample retaining only observations whose price deviates by more than 5% from its average, or store-aggregated counts) are to be stated explicitly.

Specifically: (1) identification — under what conditions on the data-generating process, including designs with low within-product price variation and dependence across products and time, is the estimand θi∗\theta_i^* identified; (2) consistency — does θ^i\hat{\theta}_i converge to θi∗\theta_i^*, pointwise in the item (as T→∞T\to\infty for a fixed item) or uniformly over items (as N→∞N\to\infty in a triangular array), and, where prices are endogenous or demand shocks hit a category simultaneously — the source evaluates its model under an assumption of independent price shocks and flags these features as challenges it does not address — the limit attained is the corresponding pseudo-true value, whose characterization belongs to the same question about the estimand rather than a separate part of the theory; (3) convergence rates — what rates are achievable as the number of observations and network capacity grow, capacity being allowed to grow with the sample size as in sieve estimation, with the admissible growth to be specified, and, because the source's variance lemma concerns a fixed design, stating whether the limit is taken at fixed positive price variance, with constants allowed to depend on it, or jointly as price variance shrinks; and (4) inference — does an asymptotic distribution exist for item-level parameters or derived elasticities (possibly at the level of smooth functionals) that supports valid confidence statements, and how can it be estimated? The theory is sought for (1)–(4) as a whole.

The question is also whether the program can fail: the negative result sought is a proof that no estimator of the stated architecture can consistently recover the item-level estimand on the data-generating processes the method claims to handle — a panel with many items, competitors' prices, and low but strictly positive within-product price variation, in which the item-level estimand is identified in principle — or an identification or estimation failure of this estimator that provably persists in the limit as the sample grows. A negative result that is not specific to the estimator is outside the scope of the question; it includes non-identification in a design with no within-product price variation at all, and the divergence of the price-coefficient variance as residualized price variance tends to zero, which the source's own Lemma proves for the regression benchmark. Here "the stated architecture" means the composition θi=FFC(Femb(Si),Xi)\theta_i=F_{FC}(F_{emb}(S_i),X_i) with the price excluded from the network inputs and least-squares training on observed sales, and such a failure must hold for every admissible network capacity, embedding dimension, and regularization within that composition.

References

  1. Kirill Safonov (2024). Neural Network Approach to Demand Estimation and Dynamic Pricing in Retail. arXiv:2412.00920. https://arxiv.org/abs/2412.00920
  2. Max H. Farrell, Tengyuan Liang, Sanjog Misra (2021). Deep Neural Networks for Estimation and Inference. Econometrica 89(1), 181-213. doi:10.3982/ECTA16901. https://doi.org/10.3982/ecta16901
  3. Xiaohong Chen, H. White (1999). Improved rates and asymptotic normality for nonparametric neural network estimators. IEEE Transactions on Information Theory 45(2), 682-691. doi:10.1109/18.749011. https://doi.org/10.1109/18.749011
  4. Xiaohong Chen (2007). Large Sample Sieve Estimation of Semi-Nonparametric Models. Handbook of Econometrics, Vol. 6B, Chapter 76, 5549-5632. doi:10.1016/S1573-4412(07)06076-X. https://doi.org/10.1016/S1573-4412(07)06076-X
  5. Steven Berry, James Levinsohn, Ariel Pakes (1995). Automobile Prices in Market Equilibrium. Econometrica 63(4), 841-890. doi:10.2307/2171802. https://doi.org/10.2307/2171802
  6. Joris Pinkse, Margaret E. Slade, Craig Brett (2002). Spatial Price Competition: A Semiparametric Approach. Econometrica 70(3), 1111-1153. doi:10.1111/1468-0262.00320. https://doi.org/10.1111/1468-0262.00320
  7. Christopher Conlon, Jeff Gortmaker (2020). Best Practices for Differentiated Products Demand Estimation with PyBLP. The RAND Journal of Economics 51(4), 1108-1161. doi:10.1111/1756-2171.12352. https://doi.org/10.1111/1756-2171.12352
  8. Marshall Fisher, Santiago Gallino, Jun Li (2018). Competition-Based Dynamic Pricing in Online Retailing: A Methodology Validated with Field Experiments. Management Science 64(6), 2496-2514. doi:10.1287/mnsc.2017.2753. https://doi.org/10.1287/mnsc.2017.2753
  9. Patrick Bajari, Denis Nekipelov, Stephen P. Ryan, Miaoyu Yang (2015). Machine Learning Methods for Demand Estimation. American Economic Review 105(5), 481-485. doi:10.1257/aer.p20151021. https://doi.org/10.1257/aer.p20151021
  10. Giovanni Compiani, Ilya Morozov, Stephan Seiler (2025). Demand Estimation with Text and Image Data. arXiv:2503.20711. https://arxiv.org/abs/2503.20711
  11. Yafei Han, Francisco Camara Pereira, Moshe Ben-Akiva, Christopher Zegras (2020). A Neural-embedded Choice Model: TasteNet-MNL Modeling Taste Heterogeneity with Flexibility and Interpretability. arXiv:2002.00922. https://arxiv.org/abs/2002.00922
  12. Ali Aouad, Antoine Désir (2022). Representing Random Utility Choice Models with Neural Networks. arXiv:2207.12877. https://arxiv.org/abs/2207.12877
  13. Haichen Hu, David Simchi-Levi (2025). Perturbing the Derivative: Doubly Wild Refitting for Model-Free Evaluation of Opaque Machine Learning Predictors. arXiv:2511.18789. https://arxiv.org/abs/2511.18789