Confidence-Weighted Return (CWR)

• Confidence-Weighted Return (CWR) is a metric that quantifies how an agent’s self-reported confidence aligns with its actual performance in financial trading and reinforcement learning.
• It integrates confidence into position sizing and reward design, systematically rewarding accurate, high-confidence predictions while penalizing overconfident errors.
• Empirical studies using PolyBench and ConfClip show that CWR effectively distinguishes well-calibrated models from those with high accuracy but misaligned confidence.

Confidence-Weighted Return (CWR) is a metric that quantifies the alignment between an agent’s self-reported confidence and actual performance, primarily developed to evaluate predictive models in financial trading and decision-making settings. Originating independently in both the prediction market benchmarking framework PolyBench and the reinforcement learning (RL) procedure ConfClip, CWR embeds confidence-calibrated risk management and reward allocation directly into quantitative model evaluation and training. This approach systematically rewards correct, high-confidence predictions while penalizing overconfidence in incorrect predictions, offering a rigorous metric for risk-sensitive evaluation and training of LLMs and RL agents (Cheng et al., 3 Apr 2026, Zhang et al., 22 Sep 2025).

1. Formal Definition and Mathematical Formulation

Prediction Markets (PolyBench)

CWR is defined over a series of model-generated predictions, each accompanied by a confidence score $c_k \in [0,1]$, and instantiated via market execution:

• Position sizing: For prediction $k$, the notional bet size is $B_k = c_k L$, where $L$ is a global base lot size (e.g., \$10).
• Order execution: The agent simulates a sweep against a historical Central Limit Order Book (CLOB), purchasing up to $B_k$ of “Yes” shares at the available ask prices. The actual capital spent is $I_k \leq B_k$, for $S_k$ shares acquired.
• Outcome resolution: After event resolution, the realized market outcome is $o_k \in \{0,1\}$.

The net profit for trade $k$ is:

$$\pi_k = o_k S_k - I_k \quad \text{(Equation 4, [2604.14199])}$$

Aggregate CWR is:

$k$0

where $k$1 is the number of executed trades (those with $k$2).

Reinforcement Learning (ConfClip)

In RL with verifiable outcomes, ConfClip defines CWR within the reward structure of episodic problems:

• Confidence estimate: For episode $k$3, the model’s confidence $k$4 is the sequence-level geometric mean probability:

$k$5

• Raw reward: For correctness reward $k$6,

$k$7

• Clipped CWR reward:

$k$8

The episodic return is $k$9 for discount $B_k = c_k L$0 (typically $B_k = c_k L$1). (Zhang et al., 22 Sep 2025)

2. Role of Confidence in Position Sizing and Policy Update

In both PolyBench and ConfClip, confidence scores directly determine the effective magnitude of model actions:

• In PolyBench, confidence $B_k = c_k L$2 scales notional trade size. High-confidence ($B_k = c_k L$3) trades receive full allocation, while lower confidences linearly downscale exposure. Thus, the influence of correct/incorrect predictions is modulated by self-asserted certainty.
• In ConfClip, reward magnitude is proportional to sequence-level confidence $B_k = c_k L$4. This coupling incentivizes the policy to express high confidence only when correct, as overconfident mistakes are maximally penalized (by negative, confidence-weighted rewards), while correct low-confidence responses receive minimal positive reinforcement. Clipping is applied to maintain gradient stability.

In both frameworks, the CWR mechanism inherently probes calibration: models that are accurate only when most confident achieve high CWR, while systematic overconfidence entails outsized losses, driving CWR negative even if raw accuracy remains above chance (Cheng et al., 3 Apr 2026, Zhang et al., 22 Sep 2025).

3. Motivations and Theoretical Rationale

CWR was introduced to address deficiencies in binary accuracy, traditional financial metrics, and coarse RL rewards:

• Prediction markets: Existing evaluation criteria (directional accuracy, plain return, or Sharpe ratio) fail to capture the economic risk of miscalibration. CWR ties model conviction to capital risk, mirroring real-world trading discipline, where capital allocation is a function of research strength and certainty (Cheng et al., 3 Apr 2026).
• Reinforcement learning: Binary verifiable rewards provide sparse and uninstructive signals. Confidence-weighted (and clipped) objectives inject calibrated granularity, supervising the model to match stated probability with actual correctness signal, thus regularizing self-assessment and discouraging speculative overconfidence (Zhang et al., 22 Sep 2025).

This design reflects established quantitative finance and Bayesian decision theory: calibrated forecasters and portfolio managers outperform those with misaligned beliefs and exposures.

4. Computation Workflows and Implementation Details

PolyBench CWR Workflow

1. Preprocessing: Select trades with confidence $B_k = c_k L$5.
2. Sizing: For each trade, compute notional exposure $B_k = c_k L$6.
3. Order simulation: Sweep CLOB up to $B_k = c_k L$7, obtaining $B_k = c_k L$8.
4. Outcome and profit: After event resolution ($B_k = c_k L$9 known), compute $L$0.
5. Aggregate: $L$1

Example Calculation: For %%%%32$k$033%%%%c_k=0.8$L$40.50, $L$5, $L$6, if $L$7 then $L$8 (100% return), if $L$9 then $B_k$0 (–100%) (Cheng et al., 3 Apr 2026).

ConfClip CWR Workflow

1. Generate responses: For each prompt $B_k$1, sample candidate answers $B_k$2.
2. Compute confidences: $B_k$3.
3. Assign reward: $B_k$4 if correct, $B_k$5 if incorrect. Compute $B_k$6.
4. Clip reward: $B_k$7 via specified $B_k$8 (e.g., $B_k$9 for correct, $I_k \leq B_k$0 for incorrect).
5. Advantage normalization: Compute group advantage $I_k \leq B_k$1 for policy gradient stabilization.
6. Policy update: Perform gradient step using $I_k \leq B_k$2-weighted REINFORCE loss plus KL-penalty (Zhang et al., 22 Sep 2025).

5. Relation to Alternative Metrics

PolyBench situates CWR within a suite of performance metrics:

Metric	Returns Adjusted For	Strengths
Confidence-Weighted Return (CWR)	None (absolute net return/dollar)	Risk-calibration, economic realism
Annualized Percentage Yield (APY)	Time (capital lockup)	Considers holding period, annualizes returns
Sharpe Ratio	Volatility	Penalizes erratic performance, risk-adjusted return

CWR is the most direct “did you make money?” criterion under live market conditions, but ignores both investment duration (APY) and variance (Sharpe). No single metric fully subsumes the others; complete assessment requires all three alongside directional accuracy (Cheng et al., 3 Apr 2026).

In RL, CWR-enhanced reward addresses sparsity and instantiates a fine-grained, calibrated learning signal; raw episodic accuracy and loss alone fail to penalize overconfident errors with sufficient granularity (Zhang et al., 22 Sep 2025).

6. Empirical Outcomes and Comparative Analysis

Prediction Markets (PolyBench)

Only MiMo-V2-Flash (17.6% CWR) and Gemini-3-Flash (6.2% CWR) achieved positive CWRs on real-time Polymarket data (~5100 trades, $I_k \leq B_k$310$, average confidence ~0.86–0.87). The remaining five LLMs reported high average confidence but posted negative CWRs, demonstrating systematic overconfidence. These results establish CWR’s utility for distinguishing between mere fluency and robust probabilistic forecasting under market uncertainty (Cheng et al., 3 Apr 2026).

Reinforcement Learning (ConfClip)

ConfClip applied to Qwen2.5-3B and Qwen2.5-7B yielded notable accuracy improvements and stabilized training across MATH, GSM8K, AIME24, and MMLU-Pro benchmarks. Ablation reveals that omitting clipping leads to collapse (100% incorrect, low-confidence outputs), while CWR with clipping ensures smooth learning dynamics, confidence-accuracy alignment, and reduced inference token consumption once the model has mastered the task (Zhang et al., 22 Sep 2025).

Key experimental results (Qwen2.5-3B, accuracy): MATH 64.34% (ConfClip) vs 30.24% (pretrained) vs 63.07% (GRPO); GSM8K 79.23% (ConfClip) vs 68.16% (pretrained); GSPO+Conf (without clip) diverges, confirming the essential role of bounded confidence-weighted rewards.

7. Significance, Limitations, and Outlook

CWR introduces an economically and decision-theoretically grounded perspective to both prediction market benchmarking and RL optimization. It operationalizes the principle that good decision-makers—human or artificial—must align capital allocation and expressed certainty with objective correctness. A model achieving high CWR is both accurate and calibrated; negative CWR in the presence of high confidence signals a critical misalignment that simple accuracy statistics cannot detect.

Limitations include time-independence (for APY one must account for capital lockup) and volatility-blindness (addressed by Sharpe ratio). In RL contexts, CWR requires stability augmentation (clipping), without which degenerate confidence patterns emerge. Across both domains, CWR provides a robust, contamination-proof yardstick for meaningful probabilistic reasoning—beyond surface-level language fluency—for LLMs and learning agents (Cheng et al., 3 Apr 2026, Zhang et al., 22 Sep 2025).