arxiv: 2603.04105 · v3 · submitted 2026-03-04 · 💰 econ.GN · q-fin.EC

Recognition: no theorem link

A Random Rule Model

Avner Seror

Authors on Pith no claims yet

Pith reviewed 2026-05-15 16:56 UTC · model grok-4.3

classification 💰 econ.GN q-fin.EC

keywords random rule modelstochastic choicedecision ruleslottery choicesmenu characteristicstwo-step identificationneural network benchmark

0 comments

The pith

Stochastic choice arises from switching among a small library of deterministic decision rules weighted by menu characteristics.

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

The paper models stochastic choice as environment-dependent switching among a small library of deterministic decision rules. Menu-level probabilities are formed by weighting named rules according to observable menu features, with a two-step identification that first uses within-feature decisive-side variation to recover relative rule weights and then uses cross-feature richness to recover the switching gate. When applied to binary lottery data, the weights concentrate on a small subset of rules and change systematically with complexity and dispersion asymmetry. The resulting model accounts for nearly all of the predictive performance of an unrestricted neural-network benchmark while remaining interpretable and passing permutation diagnostics.

Core claim

We model stochastic choice as environment-dependent switching among a small library of deterministic decision rules. A Random Rule Model generates menu-level choice probabilities via named, interpretable rules weighted by observable menu characteristics. Identification has a two-step structure: within-feature decisive-side variation identifies relative rule weights; cross-feature richness identifies the gate. Applied to binary lottery choices, the estimated weights concentrate on a small subset of rules and shift systematically with complexity and dispersion asymmetry. The model closes nearly all of the prediction gap to a flexible neural-network benchmark, while remaining interpretable, and

What carries the argument

The Random Rule Model, which forms menu-level choice probabilities by weighting a fixed library of deterministic decision rules according to observable menu characteristics, using a two-step identification that separates rule weights from the switching gate.

If this is right

Rule weights concentrate on a small subset of the available rules.
Weights shift systematically with menu complexity and dispersion asymmetry.
The model nearly closes the out-of-sample prediction gap relative to a flexible neural-network benchmark.
The model remains restrictive under permutation diagnostics and portable to independent data.

Where Pith is reading between the lines

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

The same rule-switching structure could be tested in non-lottery domains such as consumer goods or policy choices by collecting menu-feature data.
Standard behavioral models may emerge as special cases once the rule library and gate are allowed to depend on menu observables.
Direct experimental manipulation of complexity and asymmetry could test whether the predicted shifts in rule weights appear in new choices.

Load-bearing premise

That stochastic choice is generated by environment-dependent switching among a small library of deterministic decision rules and that the two-step identification using within-feature and cross-feature variation correctly recovers the rule weights and the gate.

What would settle it

Applying the estimated model to fresh binary lottery data and finding that the recovered rule weights do not concentrate on a small subset or that the model no longer accounts for most of the neural-network prediction gap.

read the original abstract

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.

Desk Editor's Note private letter to a colleague

The Random Rule Model gives an interpretable alternative to black-box fits for lottery choice that nearly matches a neural net, but the two-step identification looks vulnerable in data with few menu features.

read the letter

The paper models stochastic choice as switching among a small library of deterministic rules, with weights that depend on observable menu characteristics. On binary lottery data the weights concentrate on a few rules, shift with complexity and dispersion asymmetry, and the whole thing closes most of the gap to a flexible neural-network benchmark while staying portable to a second dataset and restrictive under permutation tests. That combination is the main new piece: a restrictive but named rule set plus explicit two-step identification that tries to separate weights from the switching gate. It is useful for anyone who wants something between rigid theory and uninterpretable ML in decision theory or mechanism design work. The empirical patterns on rule concentration and the out-of-sample portability are the parts that look cleanest. The soft spot is the identification claim. Binary lotteries have limited features, so the cross-feature variation used to identify the gate may be too thin or correlated to do the job cleanly. If the gate is not well separated, the reported weights could be contaminated and the good fit to the neural net would be less informative than it appears. The abstract gives no equations or robustness tables on this point, so it is hard to judge how much the results depend on the separation working. This is for readers in behavioral economics and decision theory who care about interpretable models of uncertain choice. It is worth a serious referee because the idea is distinct from standard random-utility or mixture setups and the performance numbers are concrete, even if the identification section will need tightening and more diagnostics.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a Random Rule Model in which stochastic choice arises from environment-dependent switching among a small library of deterministic decision rules. Menu-level choice probabilities are generated by weighting these rules according to observable menu characteristics. Identification proceeds in two steps: within-feature decisive-side variation identifies relative rule weights, while cross-feature richness identifies the gate function. Applied to binary lottery choices, the estimated weights concentrate on a small subset of rules and shift systematically with complexity and dispersion asymmetry. The model closes nearly all of the prediction gap to a flexible neural-network benchmark, remains interpretable, passes permutation diagnostics, and is portable to an independent dataset.

Significance. If the two-step identification is valid, the paper supplies an interpretable, restrictive alternative to black-box models that still achieves near-equivalent predictive performance. The concentration of weights on few rules, systematic shifts with menu features, and out-of-sample portability are potentially important contributions to modeling stochastic choice in discrete settings.

major comments (2)

[Identification] Identification section: The claim that cross-feature richness separately identifies the gate function rests on the assumption of sufficient independent variation across features. In binary lottery menus, which contain only a small number of features (probabilities and outcomes for two options), this variation may be too sparse or correlated to cleanly separate the gate from the rule weights, risking contamination of the estimated weights. This is load-bearing for the central performance claims relative to the neural-network benchmark.
[Empirical results] Empirical application: The statement that weights 'concentrate on a small subset of rules' and 'close nearly all of the prediction gap' requires explicit reporting of the exact rule library, the numerical metrics (e.g., log-likelihood or hit rates with standard errors), and the precise definition of the neural-network benchmark. Without these details, it is impossible to evaluate whether the fit reflects genuine separation or flexibility within the model class.

minor comments (1)

[Abstract] The abstract would be clearer if it named the specific rules in the library and the source of the binary lottery data.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive comments. We address each major comment below and describe the revisions we will undertake to strengthen the identification argument and empirical transparency.

read point-by-point responses

Referee: [Identification] Identification section: The claim that cross-feature richness separately identifies the gate function rests on the assumption of sufficient independent variation across features. In binary lottery menus, which contain only a small number of features (probabilities and outcomes for two options), this variation may be too sparse or correlated to cleanly separate the gate from the rule weights, risking contamination of the estimated weights. This is load-bearing for the central performance claims relative to the neural-network benchmark.

Authors: We acknowledge the importance of verifying sufficient independent variation. Although the menus involve only probabilities and payoffs, the dataset contains substantial menu-to-menu heterogeneity in these features, which we argue supplies the required cross-feature richness. To make this explicit, the revised manuscript will include Monte Carlo experiments calibrated to the empirical distribution of menu characteristics; these simulations will show that the two-step procedure recovers the gate parameters and rule weights with negligible bias. We will also report the correlation matrix and variance inflation factors among the menu features to document that multicollinearity does not prevent separation. These additions will directly address the concern while preserving the original identification logic. revision: yes
Referee: [Empirical results] Empirical application: The statement that weights 'concentrate on a small subset of rules' and 'close nearly all of the prediction gap' requires explicit reporting of the exact rule library, the numerical metrics (e.g., log-likelihood or hit rates with standard errors), and the precise definition of the neural-network benchmark. Without these details, it is impossible to evaluate whether the fit reflects genuine separation or flexibility within the model class.

Authors: We agree that greater precision is required. The revised version will (i) list every rule in the library together with its functional form, (ii) report in-sample and out-of-sample log-likelihoods and hit rates for the Random Rule Model and the neural-network benchmark, each accompanied by bootstrapped standard errors, and (iii) specify the neural-network architecture (number of hidden layers, units per layer, activation functions, regularization, and training algorithm) and the exact loss function used. These additions will allow readers to assess the magnitude of the performance gap and the degree of rule concentration directly. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper's identification relies on a two-step structure using distinct data variations (within-feature decisive-side for rule weights, cross-feature richness for the gate). No quoted equations or steps reduce predictions to fitted inputs by construction, nor involve self-definitional mappings, load-bearing self-citations, or smuggled ansatzes. The model is presented as fitted to data with out-of-sample portability checks, keeping the central claims independent of the inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The central modeling choice is the existence of a small library of deterministic rules whose weights are identified separately from the switching gate; no numerical free parameters or new entities are specified.

free parameters (1)

rule weights
Estimated from within-feature variation in the data; specific values not reported in abstract.

axioms (1)

domain assumption Stochastic choice is produced by switching among a small library of deterministic decision rules whose relative weights depend on observable menu characteristics.
Core premise stated in the abstract that enables the entire modeling and identification strategy.

pith-pipeline@v0.9.0 · 5379 in / 1413 out tokens · 59752 ms · 2026-05-15T16:56:54.768254+00:00 · methodology