Paper: arXiv 2611.22730
Authors: Paul Schneider
Abstract
We estimate the probability that an item with given covariates belongs to a set when the cross-section changes over time and items and periods are dependent. We propose two convex models whose fitted probabilities are nonnegative by construction and remain valid whatever the dependence between items, for which we establish consistency, finite-sample bounds and limit distributions for serially dependent panels. One suits events of moderate or high probability, the other rare events, and half the disagreement between their fits estimates a lower bound on the misspecification error of the worse fit, without knowledge of the true probabilities. On monthly US stock failures, the rare-event model predicts better than the ordinary logit and the classical failure logit. With the eight classical predictors, most of the gain comes from basis functions of the rank angle of each predictor and summary.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: This paper presents a highly mathematical approach to modeling inclusion probabilities on unbalanced panels, with rigorous theoretical derivations for consistency and limit distributions. It demonstrates strong empirical application to US stock failures, outperforming standard models. The novelty lies in combining sum-of-squares parametrization with quasi-likelihood for serially dependent panels, addressing key limitations of existing methods.
Research Flowchart
flowchart TD
A["Research Goal: Estimate Covariate-Indexed Inclusion Probabilities on Unbalanced Panels"] --> B{"Key Methodology: Develop Convex Models for Inclusion Probabilities"};
B -- Model 1: Moderate/High Prob. --> C1["Properties: Nonnegative, Valid for any Dependence, Consistency, Finite-Sample Bounds, Limit Distributions"];
B -- Model 2: Rare Events --> C2["Properties: Nonnegative, Valid for any Dependence, Consistency, Finite-Sample Bounds, Limit Distributions"];
C1 & C2 --> D["Computational Process: Model Fitting & Evaluation"];
D -- Data: Monthly US Stock Failures, 8 Classical Predictors --> E["Key Findings: Rare-Event Model Outperforms Ordinary/Classical Logit"];
E -- Gain attributed to --> F["Basis Functions of Rank Angle for Predictors & Summary"];
E -- Also provides --> G["Misspecification Error Lower Bound without True Probabilities"];
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes two convex models for estimating conditional marginal inclusion probabilities that are non‑negative by construction and remain valid under arbitrary dependence between items.
Source sentence
We propose two convex models whose fitted probabilities are nonnegative by construction and remain valid whatever the dependence between items, for which we establish consistency, finite-sample bounds and limit distributions for serially dependent panels.
Assumes 6
- The panel process {ξ_t} is strictly stationary with marginal law P. statistical
Source sentence
Throughout, 7 {ξt}t≥0 is strictly stationary with marginal law P, and all population objects refer to this marginal.
- The panel process {ξ_t} is ergodic. statistical
Source sentence
Let {ξt}t≥0 be strictly stationary and ergodic.
- Geometric strong mixing (or i.i.d.) is assumed for finite‑sample concentration results. statistical
Source sentence
We give almost-sure consistency of any empirical minimizer over the cone, with existence derived rather than assumed, finite-sample concentration under i.i.d. sampling and under geometric strong mixing, and a mean-squared-error bound at the parametric rate on the event of localization (Theorem 4.1).
- Positive margins of the inclusion probabilities are required (the joint law of S_{t+1} has positive margins). model
Source sentence
The theory concerns the margins (1) alone, and the identification result (Proposition 3.1) holds for any joint law with positive margins.
Show 2 more
- The covariate dictionary satisfies condition (P1) (boundedness) and the data satisfy the integrability hypothesis of Lemma A.3 on an open convex set containing the true parameter. statistical
Source sentence
Part (i) assumes the integrability hypothesis of Lemma A.3 on an open convex set O containing a point U Po ⋆ ∈Sm +. The covariate vector zt,i collects the 121 of their 153 characteristics that are missing in less than 30 percent of the name-months, rank-transformed to (−1, 1) within month, so d = 121 and the dictionary is bounded, condition (P1).
- Missing covariate entries are handled via the masked dictionary construction, assuming the weights estimate the conditional law. data
Source sentence
The masked dictionary ϕ(z) := Nref X i=1 wi(z) e(zo, xi,¯o) (7) keeps the observed coordinates at their values and averages the basis over completions of the missing ones. Appendix C.5 gives the construction, its properties and the conditions under which the weights estimate the conditional law.
Data
1963-01 – 2024-11 · US common stocks · equities · monthly
- US common‑stock universe (CRSP) market-prices public
Source sentence
7.1 Data, event, and design The panel is the US common-stock universe of Jensen et al. (2023), CRSP common stock on NYSE, AMEX and NASDAQ from January 1963 to November 2024.
- 121 characteristics from Jensen et al. (2023) fundamentals public
Source sentence
The covariate vector zt,i collects the 121 of their 153 characteristics that are missing in less than 30 percent of the name-months, rank-transformed to (−1, 1) within month, so d = 121 and the dictionary is bounded, condition (P1).
Costs
Not applicable
Finds 5
- The rare‑event model predicts monthly US stock failures better than the ordinary logit and the classical failure logit. predictive-accuracy
Source sentence
On monthly US stock failures, the rare-event model predicts better than the ordinary logit and the classical failure logit.
- Half the disagreement between the two link‑specific fits provides a lower bound on the larger of their probability‑scale errors and on the larger dictionary gap, without knowledge of the true margins. theoretical
Source sentence
One suits events of moderate or high probability, the other rare events, and half the disagreement between their fits estimates a lower bound on the misspecification error of the worse fit, without knowledge of the true probabilities. Half the disagreement between the two links bounds the larger of their errors on the probability scale from below, and through it the larger of the two dictionary gaps, with a standard
- In simulations, the composition part of the error accounts for 84 % to 99 % of the squared RMSE floor, and cannot be removed by any dictionary. statistical figure from elsewhere in the paper
Source sentence
The composition part, the root mean square of pt,i −¯p(zt,i), is 0.0186 whatever the dictionary. The dictionary part, that of ¯p−ℓPo ⋆, falls from 0.0081 at m = 4 to 0.0015 at m = 8.
- On monthly failures of US common stocks, the log‑link fit in a rank‑angle dictionary improves log‑score by about a tenth of the gain of the ordinary logit over the unconditional rate and outperforms the failure logit of Campbell et al. (2008). performance
Source sentence
On monthly failures of US common stocks the log-link fit in a rank-angle dictionary predicts better in log score and in ranking than the ordinary logit on 121 characteristics and the failure logit of Campbell et al. (2008), in log score by about a tenth of the logit’s gain over the unconditional rate and significantly in every subsample, and it stays close to the diagonal of the reliability curve in relative terms,
- The identity‑link fit has larger calibration RMSE than the logistic competitors, especially from the third decile upward. performance
Source sentence
although the logistic competitors have the smaller calibration RMSE.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 16 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-18 · same card as JSON · A language model chose these items: check the source sentence before relying on one.