pith. sign in

arxiv: 2606.06830 · v1 · pith:EIXWDKH2new · submitted 2026-06-05 · 💻 cs.CY · cs.LG

Learning Fair Demand Models

Adam N. Elmachtoub , Hyemi Kim , Jonathan Y. Tan This is my paper

Pith reviewed 2026-06-27 20:57 UTC · model grok-4.3

classification 💻 cs.CY cs.LG
keywords fair demand modelsprice optimizationdemand estimationalgorithmic fairnesssocial welfareparity fairnessRawlsian fairness
0
0 comments X

The pith

Fairness imposed on prices during demand estimation improves consumer outcomes more than fairness on demand during optimization when market sizes and prices are similar.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines how to incorporate fairness into a seller's two-stage pricing process of estimating linear demand models from customer data and then setting prices. It shows that equalizing training losses across groups can produce multiple solutions with undesirable results. For parity-wise fairness at small levels, the paper characterizes when each stage of enforcement produces higher social welfare, and finds that price fairness added in the estimation stage benefits consumers more when markets and prices are alike, while demand fairness added in optimization does better under the same condition. The two approaches coincide exactly under Rawlsian fairness.

Core claim

In the two-stage pipeline, enforcing fairness on prices in the estimation stage versus on demand in the optimization stage produces different consumer welfare outcomes under parity-wise fairness for small fairness levels, with the better choice depending on similarity of market sizes and prices in the data, while the two strategies coincide exactly under Rawlsian fairness.

What carries the argument

Two-stage pipeline of linear demand model estimation followed by price optimization, with fairness constraints applied either during estimation or during optimization.

If this is right

  • Equalizing training loss across consumer groups can result in multiple solutions and undesirable outcomes.
  • For small parity-wise fairness levels, the strategy yielding higher social welfare depends on market conditions.
  • When market sizes and prices are similar, imposing price fairness in estimation benefits consumers more.
  • Imposing demand fairness in optimization yields better consumer outcomes under similar market conditions.
  • The estimation-stage and optimization-stage strategies coincide exactly for Rawlsian fairness.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Practitioners could check observable similarity of market sizes and prices in their data to decide which stage to enforce fairness at.
  • The exact coincidence under Rawlsian fairness may extend to the alternate demand functions studied in the paper.
  • The vaccine pricing case study provides a concrete setting where the similarity condition could be measured and the welfare difference tested.

Load-bearing premise

The seller uses a stylized two-stage pipeline with linear demand model estimation followed by price optimization.

What would settle it

A dataset with similar market sizes and prices where imposing price fairness during estimation does not produce higher consumer welfare than imposing demand fairness during optimization for small parity fairness levels.

Figures

Figures reproduced from arXiv: 2606.06830 by Adam N. Elmachtoub, Hyemi Kim, Jonathan Y. Tan.

Figure 4
Figure 4. Figure 4: Under Rawlsian price fairness, increasing [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
read the original abstract

Data-driven pricing is increasingly prevalent in sectors such as airlines, lending, insurance, and retail. By learning demand models from customer features and setting prices accordingly, these systems may generate discriminatory outcomes that raise fairness concerns. This leads to fundamental questions - how and where should systems incorporate fairness considerations in the pricing pipeline, and how does it ultimately affect societal outcomes? To answer these, we study a stylized model where a seller has a two-stage decision pipeline comprising linear demand model estimation followed by price optimization. The seller considers fairness notions in training loss, price, and demand, under both parity-wise and Rawlsian perspectives. We show that equalizing training loss across consumer groups leads to multiple solutions, which in turn can result in undesirable outcomes despite being a standard approach in fair machine learning. Focusing instead on fairness applied directly to prices or demand, we compare two strategies that enforce fairness in either the demand estimation stage or the price optimization stage. For parity-wise fairness, we characterize when each strategy yields higher social welfare under small fairness levels. We show that when market sizes and prices in the dataset are similar, imposing price fairness in the estimation stage is more beneficial to consumers, whereas imposing demand fairness in the optimization stage yields better consumer outcomes. For Rawlsian fairness, the two strategies coincide exactly. Lastly, we extend our model to alternate demand functions and conduct a case study using real-world vaccine pricing data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper studies fairness in data-driven pricing via a stylized two-stage pipeline of linear demand estimation followed by price optimization. It examines parity-wise and Rawlsian fairness notions applied to training loss, prices, or demand. Equalizing training loss across groups is shown to admit multiple solutions that can produce undesirable outcomes. The core contribution compares two enforcement strategies (fairness in estimation stage vs. optimization stage): for parity-wise fairness and small fairness levels, it characterizes conditions under which each yields higher social welfare, with the result that similar market sizes and prices favor price fairness in estimation for consumers while demand fairness in optimization is better otherwise; for Rawlsian fairness the two strategies coincide exactly. The analysis is extended to alternate demand functions and illustrated with a vaccine pricing case study.

Significance. If the characterizations hold under the stated assumptions, the paper offers precise, model-based guidance on where to locate fairness constraints inside a pricing pipeline, which is directly relevant to deployed systems in airlines, lending, and retail. The exact coincidence result for Rawlsian fairness and the welfare ordering for small fairness levels under parity are clean theoretical contributions that could inform both algorithm design and regulatory discussion. The extension beyond linear demand and the real-data case study strengthen the practical relevance.

major comments (2)
  1. [Training-loss fairness analysis] The section analyzing equalizing training loss: the claim that this approach 'leads to multiple solutions, which in turn can result in undesirable outcomes' is used to motivate shifting focus to price/demand fairness, yet the manuscript does not appear to supply an explicit characterization or example quantifying the welfare loss relative to the other strategies; this weakens the justification for discarding the training-loss route.
  2. [Parity-wise fairness results] Parity-wise fairness characterization (small fairness levels): the welfare comparison between price fairness in estimation and demand fairness in optimization is stated to depend on 'market sizes and prices in the dataset are similar,' but the precise metric or threshold for similarity is not formalized (e.g., via a bound on the ratio of group sizes or price dispersion); without this, the condition remains informal and the result's applicability is hard to assess.
minor comments (2)
  1. [Model setup and notation] Notation for the fairness level parameter and the two enforcement strategies should be introduced once with consistent symbols and then used uniformly; occasional redefinition makes the comparison between stages harder to follow.
  2. [Case study] The case study section would benefit from an explicit statement of how the real-world vaccine pricing dataset was pre-processed to fit the linear demand assumption and from a sensitivity table showing welfare rankings for fairness levels beyond the 'small' regime.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and the recommendation for minor revision. We address the two major comments point by point below.

read point-by-point responses

  1. Referee: [Training-loss fairness analysis] The section analyzing equalizing training loss: the claim that this approach 'leads to multiple solutions, which in turn can result in undesirable outcomes' is used to motivate shifting focus to price/demand fairness, yet the manuscript does not appear to supply an explicit characterization or example quantifying the welfare loss relative to the other strategies; this weakens the justification for discarding the training-loss route.

    Authors: We agree that an explicit quantitative example would strengthen the motivation. The manuscript establishes that training-loss fairness admits multiple solutions but does not quantify the welfare loss relative to price- or demand-fairness approaches. In the revision we will add a short numerical example (using the linear demand model) that illustrates how different solutions under training-loss fairness can produce strictly lower social welfare than the other two routes. revision: yes

  2. Referee: [Parity-wise fairness results] Parity-wise fairness characterization (small fairness levels): the welfare comparison between price fairness in estimation and demand fairness in optimization is stated to depend on 'market sizes and prices in the dataset are similar,' but the precise metric or threshold for similarity is not formalized (e.g., via a bound on the ratio of group sizes or price dispersion); without this, the condition remains informal and the result's applicability is hard to assess.

    Authors: The welfare ordering is obtained by comparing the first-order welfare expressions under small fairness perturbations; the sign of the difference depends on the relative market sizes and price levels appearing in those expressions. While the paper states the condition in terms of similarity, it does not supply explicit quantitative bounds. We will revise the relevant theorem statement and surrounding text to include a formal similarity condition (e.g., a bound on the ratio of group sizes together with a bound on price dispersion derived from the demand parameters) so that the region of applicability is stated precisely. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The central results are explicit characterizations of welfare comparisons (for small fairness levels) and an exact coincidence result between two fairness-enforcement strategies inside a linear-demand two-stage pipeline. These follow directly from the stated model assumptions, fairness definitions, and optimization objectives without any reduction to fitted parameters by construction, self-citation chains, or renamed inputs. The analysis is parameter-free in its theoretical claims and rests on transparent algebraic comparisons rather than empirical fitting or imported uniqueness theorems.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claims rest on a stylized linear demand model and assumptions about market sizes, prices, and small fairness levels; full details unavailable from abstract alone.

free parameters (1)
  • fairness level
    Characterizations are given for small fairness levels; exact parameterization not specified in abstract.
axioms (1)
  • domain assumption Demand follows a linear model in the stylized pipeline
    Stated as the basis for the two-stage estimation-then-optimization setup.

pith-pipeline@v0.9.1-grok · 5777 in / 1210 out tokens · 18501 ms · 2026-06-27T20:57:49.914594+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 9 canonical work pages · 2 internal anchors

  1. [1]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    Social bias meets data bias: The impacts of labeling and measurement errors on fairness criteria , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  2. [2]

    Operations research , volume=

    The price of fairness , author=. Operations research , volume=. 2011 , publisher=

    2011

  3. [3]

    ACM Journal on Responsible Computing , year=

    Against predictive optimization: On the legitimacy of decision-making algorithms that optimize predictive accuracy , author=. ACM Journal on Responsible Computing , year=

  4. [4]

    International Conference on Machine Learning , pages=

    Delayed impact of fair machine learning , author=. International Conference on Machine Learning , pages=. 2018 , organization=

    2018

  5. [5]

    The Journal of Machine Learning Research , volume=

    The measure and mismeasure of fairness , author=. The Journal of Machine Learning Research , volume=. 2023 , publisher=

    2023

  6. [6]

    Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=

    Model multiplicity: Opportunities, concerns, and solutions , author=. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2022

  7. [7]

    International Conference on Machine Learning , pages=

    Predictive multiplicity in classification , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Rashomon capacity: A metric for predictive multiplicity in classification , author=. Advances in Neural Information Processing Systems , volume=

  9. [9]

    International Conference on Machine Learning , pages=

    Characterizing fairness over the set of good models under selective labels , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  10. [10]

    Proceedings of the 2021 ACM conference on fairness, accountability, and transparency , pages=

    Fairness, welfare, and equity in personalized pricing , author=. Proceedings of the 2021 ACM conference on fairness, accountability, and transparency , pages=

    2021

  11. [11]

    Proceedings of the ACM Web Conference 2022 , pages=

    Regulatory instruments for fair personalized pricing , author=. Proceedings of the ACM Web Conference 2022 , pages=

    2022

  12. [12]

    Management Science , volume=

    Price discrimination with fairness constraints , author=. Management Science , volume=. 2022 , publisher=

    2022

  13. [13]

    Operations Research , year=

    Dynamic pricing with fairness constraints , author=. Operations Research , year=

  14. [14]

    Operations Research , year=

    Fair fares for vehicle sharing systems , author=. Operations Research , year=

  15. [15]

    Management Science , year=

    Utility fairness in contextual dynamic pricing with demand learning , author=. Management Science , year=

  16. [16]

    Dynamic Pricing with Demand Covariates

    Dynamic pricing with demand covariates , author=. arXiv preprint arXiv:1604.07463 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    International Conference on Machine Learning , pages=

    Pricing experimental design: causal effect, expected revenue and tail risk , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  18. [18]

    Operations Research , volume=

    Nonparametric pricing analytics with customer covariates , author=. Operations Research , volume=. 2021 , publisher=

    2021

  19. [19]

    Management Science , volume=

    Personalized dynamic pricing with machine learning: High-dimensional features and heterogeneous elasticity , author=. Management Science , volume=. 2021 , publisher=

    2021

  20. [20]

    PloS one , volume=

    The willingness to pay for vaccination against tick-borne encephalitis and implications for public health policy: evidence from Sweden , author=. PloS one , volume=. 2015 , publisher=

    2015

  21. [21]

    Statistical science , volume=

    Statistical modeling: The two cultures (with comments and a rejoinder by the author) , author=. Statistical science , volume=. 2001 , publisher=

    2001

  22. [22]

    arXiv preprint arXiv:2405.19667 , year=

    Reconciling Model Multiplicity for Downstream Decision Making , author=. arXiv preprint arXiv:2405.19667 , year=

    work page arXiv

  23. [23]

    Sociological Methods & Research , volume=

    Fairness in criminal justice risk assessments: The state of the art , author=. Sociological Methods & Research , volume=. 2021 , publisher=

    2021

  24. [24]

    Applied ethics , pages=

    A theory of justice , author=. Applied ethics , pages=. 2017 , publisher=

    2017

  25. [25]

    Proceedings of the 3rd innovations in theoretical computer science conference , pages=

    Fairness through awareness , author=. Proceedings of the 3rd innovations in theoretical computer science conference , pages=

  26. [26]

    Big data , volume=

    Fair prediction with disparate impact: A study of bias in recidivism prediction instruments , author=. Big data , volume=. 2017 , publisher=

    2017

  27. [27]

    Advances in neural information processing systems , volume=

    Equality of opportunity in supervised learning , author=. Advances in neural information processing systems , volume=

  28. [28]

    Inherent Trade-Offs in the Fair Determination of Risk Scores

    Inherent trade-offs in the fair determination of risk scores , author=. arXiv preprint arXiv:1609.05807 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    Advances in neural information processing systems , volume=

    Counterfactual fairness , author=. Advances in neural information processing systems , volume=

  30. [30]

    Proceedings of the 23rd acm sigkdd international conference on knowledge discovery and data mining , pages=

    Algorithmic decision making and the cost of fairness , author=. Proceedings of the 23rd acm sigkdd international conference on knowledge discovery and data mining , pages=

  31. [31]

    Management Science , year=

    Learning to Be Fair: A Consequentialist Approach to Equitable Decision Making , author=. Management Science , year=

  32. [32]

    Proceedings of the 2020 conference on fairness, accountability, and transparency , pages=

    Fairness and utilization in allocating resources with uncertain demand , author=. Proceedings of the 2020 conference on fairness, accountability, and transparency , pages=

    2020

  33. [33]

    arXiv preprint arXiv:2406.13882 , year=

    Allocation requires prediction only if inequality is low , author=. arXiv preprint arXiv:2406.13882 , year=

    work page arXiv

  34. [34]

    ACM Journal on Responsible Computing , volume=

    Against predictive optimization: On the legitimacy of decision-making algorithms that optimize predictive accuracy , author=. ACM Journal on Responsible Computing , volume=. 2024 , publisher=

    2024

  35. [35]

    Proceedings of the 2020 conference on fairness, accountability, and transparency , pages=

    Mitigating bias in algorithmic hiring: Evaluating claims and practices , author=. Proceedings of the 2020 conference on fairness, accountability, and transparency , pages=

    2020

  36. [36]

    The Journal of Finance , volume=

    Predictably unequal? The effects of machine learning on credit markets , author=. The Journal of Finance , volume=. 2022 , publisher=

    2022

  37. [37]

    International Conference on Machine Learning , pages=

    Loss balancing for fair supervised learning , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  38. [38]

    2021 , publisher=

    Pyomo--optimization modeling in python , author=. 2021 , publisher=

    2021

  39. [39]

    and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and

    Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E. and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and. Nature Methods , year =

  40. [40]

    2012 , publisher=

    Matrix analysis , author=. 2012 , publisher=

    2012

  41. [41]

    Production and Operations Management , pages=

    EXPRESS: Data-Driven Price Discrimination, Regulation, and Platform Compliance: Evidence from a Field Experiment and a Natural Experiment , author=. Production and Operations Management , pages=

  42. [42]

    ACM computing surveys (CSUR) , volume=

    A survey on bias and fairness in machine learning , author=. ACM computing surveys (CSUR) , volume=. 2021 , publisher=

    2021

  43. [43]

    Operations Research , year=

    Fairness-aware online price discrimination with nonparametric demand models , author=. Operations Research , year=

  44. [44]

    Ai & Society , volume=

    On prediction-modelers and decision-makers: why fairness requires more than a fair prediction model , author=. Ai & Society , volume=. 2025 , publisher=

    2025

  45. [45]

    The quarterly journal of economics , volume=

    Human decisions and machine predictions , author=. The quarterly journal of economics , volume=. 2018 , publisher=

    2018

  46. [46]

    Annals of Operations Research , volume=

    A guide to formulating fairness in an optimization model , author=. Annals of Operations Research , volume=. 2023 , publisher=

    2023

  47. [47]

    Operations Research , volume=

    Best of both worlds: Ex ante and ex post fairness in resource allocation , author=. Operations Research , volume=. 2024 , publisher=

    2024

  48. [48]

    Operations Research , year=

    Algorithmic challenges in ensuring fairness at the time of decision , author=. Operations Research , year=

  49. [49]

    Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency , pages=

    Operationalizing the search for less discriminatory alternatives in fair lending , author=. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2024

  50. [50]

    Human Vaccines & Immunotherapeutics , volume=

    Cost-based COVID-19 vaccination and willingness to pay: A post-pandemic review , author=. Human Vaccines & Immunotherapeutics , volume=. 2024 , publisher=

    2024

  51. [51]

    Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=

    On the fairness of machine-assisted human decisions , author=. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2022

  52. [52]

    The University of Chicago Law Review , volume=

    Big data and discrimination , author=. The University of Chicago Law Review , volume=. 2019 , publisher=

    2019

  53. [53]

    Management Science , volume=

    Assessing algorithmic fairness with unobserved protected class using data combination , author=. Management Science , volume=. 2022 , publisher=

    2022

  54. [54]

    Management Science , volume=

    On the efficiency-fairness trade-off , author=. Management Science , volume=. 2012 , publisher=

    2012

  55. [55]

    arXiv preprint arXiv:2301.10642 , year=

    Group fairness in dynamic refugee assignment , author=. arXiv preprint arXiv:2301.10642 , year=

    work page arXiv

  56. [56]

    arXiv preprint arXiv:2208.07341 , year=

    Fair assortment planning , author=. arXiv preprint arXiv:2208.07341 , year=

    work page arXiv

  57. [57]

    arXiv preprint arXiv:2311.13765 , year=

    Learning optimal and fair policies for online allocation of scarce societal resources from data collected in deployment , author=. arXiv preprint arXiv:2311.13765 , year=

    work page arXiv

  58. [58]

    Operations Research , volume=

    A unified framework for analyzing and optimizing a class of convex fairness measures , author=. Operations Research , volume=. 2026 , publisher=

    2026

  59. [59]

    arXiv preprint arXiv:2507.05606 , year=

    Fairness-Aware Static and Dynamic Assortment Optimization: Optimal Selection with Balanced Market Share , author=. arXiv preprint arXiv:2507.05606 , year=

    work page arXiv

  60. [60]

    2023 , publisher=

    A simple way towards fair assortment planning: Algorithms and welfare implications , author=. 2023 , publisher=

    2023

  61. [61]

    ACM Transactions on Economics and Computation , volume=

    Fairness maximization among offline agents in online-matching markets , author=. ACM Transactions on Economics and Computation , volume=. 2023 , publisher=

    2023

  62. [62]

    arXiv preprint arXiv:2011.13908 , year=

    Group-level fairness maximization in online bipartite matching , author=. arXiv preprint arXiv:2011.13908 , year=

    work page arXiv 2011