Testing by Betting while Borrowing and Bargaining

Basic Information

• Paper ID: 2407.11465
• Title: Testing by Betting while Borrowing and Bargaining
• Authors: Hongjian Wang, Aaditya Ramdas (Carnegie Mellon University)
• Classification: math.ST math.PR q-fin.MF stat.ME stat.TH
• Publication Date: October 17, 2025 (arXiv v2)
• Paper Link: https://arxiv.org/abs/2407.11465

Abstract

This paper investigates an extension of the "testing by betting" framework in game-theoretic statistics. In the traditional framework, betting scores (or more generally, e-processes) are used to quantify evidence against null hypotheses. A core assumption in prior work is that bettors cannot wager more than their current wealth. This paper explores what happens when bettors are allowed to borrow money after bankruptcy, proposing various (adjusted) definitions of evidence relative to borrowing, debt, and accumulated wealth. It also examines scenarios where bettors can "bargain" to obtain better odds than those specified by the null hypothesis.

Research Background and Motivation

Problem Definition

The core problem addressed is: in the game-theoretic framework of hypothesis testing, how should statistical evidence be redefined and quantified when the traditional constraint of "not wagering more than current wealth" is relaxed?

Research Significance

1. Theoretical Extension: The traditional testing-by-betting framework assumes non-negative bettor wealth, a restriction that is overly stringent in practice
2. Practical Application: Real-world investors and traders frequently use leverage and borrowing to expand investment scale
3. Statistical Innovation: Exploring new methods for evidence quantification may provide fresh perspectives for statistical inference

Limitations of Existing Approaches

1. Wealth Constraints: In classical work by Shafer and Vovk, bettors must exit the game immediately upon bankruptcy
2. Evidence Definition: Existing framework definitions of evidence depend on non-negativity of wealth processes
3. Lack of Flexibility: Unable to handle borrowing and leverage operations common in practice

Research Motivation

The paper is inspired by the behavior of aggressive gamblers and risk-seeking traders in real financial markets, attempting to introduce these concepts into statistical hypothesis testing and provide a more flexible theoretical framework for game-theoretic statistics.

Core Contributions

1. Extended the Classical Framework: Proposed NSM (non-negative supermartingale) betting definitions allowing borrowing (Definition 2.3)
2. Constructed New Evidence Definitions:
   • Concept of tail evidence (Definition 3.1)
   • Concept of sequential tail evidence (Definition 3.2)
3. Provided Multiple Evidence Quantification Methods:
   • Evidence based on total wealth (Section 4)
   • Evidence based on net wealth (Section 5)
   • Current evidence (Section 6)
4. Established Leverage Invariance Theory: Proved that under specific standardization, borrowing does not improve the expected value of evidence (Section 8)
5. Extended to Bargaining Scenarios: Addressed cases where betting odds can be negotiated (Section 9)

Methodology Details

Task Definition

Input: Sequence of random variables $\{X_t\}_{t≥1}$, betting strategy $\{λ_t\}_{t≥1}$, borrowing strategy $\{β_t\}_{t≥1}$Output: Statistics quantifying evidence against the null hypothesis Constraints: $β_t ≥ -W_{t-1}$ (cannot borrow more than the negative of current wealth)

Model Architecture

1. Classical Betting Game (No Borrowing)

Wealth process defined as: $W_0 = 1, \quad W_t = W_{t-1} \cdot (1 + λ_t X_t)$

where $λ_t$ is the betting proportion and $X_t ∈ \{-1, 1\}$ is the coin flip outcome.

2. Borrowing Betting Game

Wealth Process: $W_t = (W_{t-1} + β_t) \cdot B_t$

Debt Process: $L_t = \sum_{i=1}^t β_i$

Net Wealth Process: $N_t = W_t - L_t$

3. Evidence Definition Framework

(a,b)-Tail Evidence: Random variable $E$ satisfying $P(E ≥ x) ≤ \frac{a}{x-b}, \quad \forall x > b$

(a,b)-Sequential Tail Evidence: Process $\{E_t\}$ satisfying $P(\sup_t E_t ≥ x) ≤ \frac{a}{x-b}, \quad \forall x > b$

Technical Innovations

1. Doob Decomposition Extension

Proposition 2.4: The net wealth process $\{N_t\}$ is a martingale under $P_M$ and a supermartingale under $P_S$. Furthermore, under $P_M$, $W_t = N_t + L_t$ is the Doob decomposition of $\{W_t\}$.

2. Current Evidence Construction

Theorem 6.1: Define instantaneous leverage ratio $ρ_t = (W_{t-1} + β_t)/W_{t-1}$, then the process $V_t = \frac{W_t}{ρ_1 \cdots ρ_t}$ is a non-negative supermartingale under $P_S$ and a martingale under $P_M$.

3. Leverage Invariance

Proposition 8.2: The function $E(Y) = \sup_{a,b}\{E_Q(aY + b) : P(aY + b ≥ x) ≤ x^{-1}\}$ is invariant to leverage operations.

Experimental Setup

Theoretical Verification

This paper is primarily theoretical research, with verification through mathematical proofs of various propositions and theorems. Main verification approaches include:

1. Constructive Proofs: Proving existence by explicitly constructing processes satisfying the conditions
2. Counterexample Analysis: Using specific examples to illustrate limitations of certain properties
3. Numerical Examples: Such as the two-round betting example in Example 5.7

Example Analysis

Example 5.7: Consider two rounds of coin betting with $X_1, X_2 ∈ \{-1,1\}$, borrowing $β_1 = β_2 = 1$, and betting proportions $λ_1 = λ_2 = 1/2$.

Define sub-debt $\tilde{L}_2 = β_1 + (2-X_1)β_2$. Results show that in certain cases, e-values based on sub-net wealth may be larger.

Experimental Results

Main Theoretical Results

1. Total Wealth as Evidence

Proposition 4.3:

• Under bounded expected stopping time debt assumption, $W_τ$ is $(1+L, 0)$-tail evidence
• Under bounded expected debt assumption, $\{W_t\}$ is $(1+L, 0)$-sequential tail evidence

2. Net Wealth as Evidence

Proposition 5.3:

• Under net wealth lower bound assumption, $N_τ$ is $(1-N_{min}, N_{min})$-tail evidence
• $\{N_t\}$ is $(1-N_{min}, N_{min})$-sequential tail evidence

3. Leverage Invariance Results

Proposition 8.2: Proves that under appropriate standardization, borrowing does not improve the expected value of evidence, meaning borrowing fundamentally cannot obtain "better" evidence merely by changing the betting structure.

Key Findings

1. Evidence Discount: Evidence generated by borrowing must be adjusted according to the borrowing amount
2. Current Evidence Advantage: Current evidence (Theorem 6.1) does not require information about other possible worlds
3. Net Wealth Advantage: Compared to total wealth, net wealth as evidence is easier to verify and satisfies assumption conditions
4. Leverage Limitations: Theoretically proves that borrowing alone cannot improve the expected value of standardized evidence

Related Work

Game-Theoretic Statistics Foundations

1. Shafer & Vovk (2005, 2019): Established the classical testing-by-betting framework
2. Ville (1939): Provided foundational inequalities for non-negative supermartingales
3. Ramdas et al. (2023): Modern survey of game-theoretic statistics

Almost Supermartingale Theory

Robbins & Siegmund (1971): The concept of almost supermartingales provides theoretical foundation for this paper's extensions

Financial Mathematics Connections

The paper connects statistical hypothesis testing with arbitrage, interest rates, and numeraire-adjusted pricing in financial markets, demonstrating the intersection of statistics and financial mathematics.

Conclusions and Discussion

Main Conclusions

1. Feasibility: Under appropriate assumptions, valid statistical evidence can indeed be constructed in the borrowing framework
2. Diversity: Multiple evidence definition methods are provided based on total wealth, net wealth, etc.
3. Limitations: The leverage invariance theorem shows that borrowing alone theoretically does not improve evidence quality
4. Practicality: The current evidence method provides a practically implementable evidence quantification scheme

Limitations

1. Assumption Dependence: Most results depend on boundedness assumptions for borrowing amounts
2. Standardization Issues: Leverage invariance holds only under specific standardization
3. Theory-Practice Gap: Considerable distance remains between the theoretical framework and practical statistical applications
4. Computational Complexity: Computing certain evidence requires considering counterfactual scenarios

Future Directions

The paper proposes three main future research directions:

1. Sequential Evidence Improvement: Seek sequential evidence definitions depending only on current debt rather than upper bounds
2. Borrowing Advantage Conditions: Investigate when borrowing can bring practical benefits from an evidence perspective
3. Utility Function Extension: Explore alternative concave utility functions beyond logarithmic utility to accommodate negative values

In-Depth Evaluation

Strengths

1. Theoretical Innovation: First systematic introduction of borrowing concepts into the testing-by-betting framework
2. Mathematical Rigor: All major results have complete mathematical proofs
3. Multi-Perspective Analysis: Analyzes the problem from multiple angles including total wealth, net wealth, and current evidence
4. Practical Inspiration: Leverage invariance results provide profound theoretical insights
5. Complete Framework: Forms a complete theoretical system from basic definitions to advanced extensions

Weaknesses

1. Limited Practical Utility: Insufficient connection between theoretical results and practical statistical applications
2. Strong Assumptions: Most results require relatively strong technical assumptions (e.g., bounded expected debt)
3. Computational Complexity: Some evidence definitions involve counterfactual calculations, making practical implementation difficult
4. Lack of Empirical Validation: Primarily theoretical research with insufficient empirical validation on real data
5. Insufficient Application Guidance: Lacks practical guidance on when borrowing strategies should be employed

Impact

1. Theoretical Contribution: Provides important theoretical extension to game-theoretic statistics
2. Interdisciplinary Value: Connects statistics, probability theory, and financial mathematics
3. Inspirational Significance: Leverage invariance results may inspire broader statistical theory research
4. Methodological Value: The evidence definition framework provided has general methodological value

Applicable Scenarios

1. Theoretical Research: Suitable for theoretical researchers in game-theoretic statistics
2. Financial Statistics: Potentially applicable to statistical inference problems in financial markets
3. Sequential Analysis: Provides new theoretical tools for sequential hypothesis testing
4. Risk Management: May have application value in risk assessment considering leverage operations

References

Main references in this paper include:

• Shafer, G. & Vovk, V. (2005, 2019): Classical works on game-theoretic statistics
• Robbins, H. & Siegmund, D. (1971): Foundational work on almost supermartingale theory
• Ramdas, A. et al. (2023): Modern survey of game-theoretic statistics
• Ville, J. (1939): Classical inequalities in probability theory

Overall Assessment: This is a high-quality theoretical statistics paper that proposes innovative extensions in the field of game-theoretic statistics. While its practical utility requires further development, its theoretical contributions and interdisciplinary value make it an important advance in the field. The leverage invariance theorem is a particularly valuable theoretical insight that may influence broader statistical theory research.