--- title: Kelly Betting as Bayesian Model Evaluation url: https://www.emergentmind.com/topics/kelly-betting-as-bayesian-model-evaluation type: topic --- # Kelly Betting as Bayesian Model Evaluation Kelly betting as Bayesian model evaluation provides a mathematically rigorous framework for real-time, sequential assessment of probabilistic forecasting models. By treating each model or agent as a Kelly bettor and interpreting their evolving bankrolls as Bayesian credibilities, this approach unifies prediction-market dynamics, strictly proper scoring, information-theoretic optimality, and Bayesian model averaging. It yields analytic connections between log-loss, Kullback-Leibler divergence, and the rate at which the best model can be distinguished from suboptimal alternatives, while also supporting online updating and market consensus formation [1201.6655][2602.09982]. ## 1. Mathematical Foundations and Setup Consider $K$ competing models $M_1, \dots, M_K$ forecasting a sequence of binary outcomes $o_t \in \{0,1\}$ for $t = 1, \ldots, T$. Each model $i$ outputs an updated predictive probability $p_{i,t} = P_{M_i}(o_t = 1 \mid \text{history})$. Each model is assigned a bankroll (credibility) $B_{i,t}$, initialized to the prior $B_{i,0}=P(M_i)$ with normalization $\sum_i B_{i,0}=1$ [2602.09982]. At each round, models bet as Kelly agents against a “market” consensus probability $q_t$, with bets fractions given by the classical Kelly formula: \[ b_{i,t} = \frac{p_{i,t} - q_t}{1 - q_t}. \] Once the outcome $o_t$ is revealed, each model’s bankroll is updated multiplicatively: \[ B_{i,t} = B_{i,t-1} \begin{cases} 1 + b_{i,t} \bigl(1/q_t - 1\bigr), & o_t=1 \\ 1 - b_{i,t}, & o_t=0 \end{cases} \] or, equivalently, \[ B_{i,t} = B_{i,t-1} \left(\frac{p_{i,t}}{q_t}\right)^{o_t} \left(\frac{1 - p_{i,t}}{1 - q_t}\right)^{1-o_t} \] [1201.6655][2602.09982]. ## 2. Equivalence to Bayesian Model Evaluation The growth of $B_{i,t}$ implements exact Bayesian updating for model credibility: \[ B_{i,t} \propto P(M_i) \prod_{s=1}^t P(o_s | M_i). \] This alignment is seen by noting $P(o_t|M_i) = p_{i,t}^{o_t}(1-p_{i,t})^{1-o_t}$ and that the “market” aggregates model forecasts into $q_t = \sum_i B_{i, t-1}\, p_{i,t}$. The ratio $B_{i,t}/B_{j,t}$ exactly matches the posterior odds \[ \frac{B_{i,t}}{B_{j,t}} = \frac{P(M_i | o_{1:t})}{P(M_j | o_{1:t})}. \] This shows that Kelly betting yields the same sequential model evidence as Bayesian filtering, with bankrolls serving as normalized posterior credibilities at every time step [1201.6655][2602.09982]. ## 3. Market Aggregation, Log-Loss, and Regret At equilibrium, market price is the consensus forecast \[ q_t = \sum_{i=1}^K B_{i,t-1} p_{i,t}, \] with $B_{i,t-1}$ interpreted as normalized model credibilities. The incremental log-growth for model $i$ satisfies: \[ \Delta \ln B_{i,t} = o_t \ln\frac{p_{i,t}}{q_t} + (1-o_t)\ln\frac{1-p_{i,t}}{1-q_t} = -[\ell(o_t, p_{i,t}) - \ell(o_t, q_t)] \] where $\ell(o, p) = -[o\ln p + (1-o)\ln(1-p)]$ is the log-loss. Thus, maximizing log-bankroll aligns with minimizing cumulative log-loss against the market mixture [2602.09982]. The expected excess growth rate is given by the negative KL divergence between the true data-generating distribution $p^*$ and model $p$: \[ \mathbb{E}_{o \sim p^*} \left[\ln \frac{p}{p^*}\right] = -D_{\text{KL}}(p^* || p). \] A worst-case log regret bound follows via wealth conservation in prediction markets: after $T$ rounds, \[ L \leq L_i + \ln\frac{1}{w_{i,0}}, \] where $L$ is the market log-loss, $L_i$ is that for agent $i$, and $w_{i,0}$ is the prior wealth. This is the Bayesian model-evidence penalty term for expert $i$ [1201.6655]. ## 4. Posterior Evolution and Beta-Binomial Dynamics Given a data stream $y_1, ..., y_T \iid \operatorname{Bernoulli}(\pi)$, assign initial wealth $w_{i,0}$ and beliefs $p_{i,0}$. The market price sequence $p_t$ updates exactly as the posterior mean of a Beta-Binomial model: \[ \begin{aligned} \alpha_0 &= \sum_i w_{i,0} p_{i,0},\\ \beta_0 &= \sum_i w_{i,0}(1-p_{i,0}), \\ \alpha_{t+1} &= \alpha_t + x_t, \\ \beta_{t+1} &= \beta_t + (1-x_t), \\ p_{t+1} &= \frac{\alpha_{t+1}}{\alpha_{t+1} + \beta_{t+1}}. \end{aligned} \] Thus, the market price acts as the posterior predictive mean, reflecting aggregate learning as in Bayesian inference [1201.6655]. ## 5. Fractional Kelly, Tempered Posteriors, and Discounting Fractional Kelly betting generalizes the approach by scaling bet size to a confidence parameter $\lambda \in (0,1)$: \[ f_{i,t}^{(\lambda)} = \lambda\,\frac{p_{i,t} - p_t}{1 - p_t}. \] A bettor using fractional Kelly acts as a full Kelly bettor on the tempered belief $\tilde p_{i,t} = \lambda p_{i,t} + (1 - \lambda) p_t$. The consensus price becomes \[ p_t = \frac{\sum_i \lambda_i w_{i,t} p_{i,t}}{\sum_i \lambda_i w_{i,t}}. \] When all $\lambda_i \equiv \lambda$, this implements a market tracking a discounted Bernoulli process, with the price converging to a time-discounted frequency. Empirically, this yields $p_t \approx d_t$ where \[ d_t = \frac{\sum_{s=1}^t \gamma^{t-s} y_s}{\sum_{s=1}^t \gamma^{t-s}},\qquad \gamma \approx \frac{1-\lambda}{1+\lambda}. \] This provides a probabilistic interpretation for fractional Kelly betting and ties it to credibility discounting [1201.6655]. ## 6. Empirical Performance and Metric Comparison In simulation studies involving binary outcome sequences (e.g., “volleyball" matches to 100 points), Kelly-Bayes evaluation is compared to log-loss and Brier score for the task of model selection: - When the alternative model uses an incorrect but fixed $p$, Kelly selects the true model more often than log-loss/Brier (e.g., 55% vs 50%). - For alternatives with recency bias, Kelly achieves substantially higher model-picking accuracy (96% vs 73% for log-loss). - For alternatives with random drift, Kelly outperforms log-loss (74% vs 58%). Over repeated matches, when the ending bankroll is carried over as prior, Kelly-Bayes quickly outperforms and dominates these classical metrics in terms of model-selection accuracy [2602.09982]. ## 7. Real-Time Implementation and Generalizations At each time step, the market consensus is computed as the bankroll-weighted average of model forecasts: \[ q_t = \sum_i B_{i,t-1} p_{i,t}. \] Bankroll is updated as above, and normalization ensures $B_{i,t}$ retains the interpretation as posterior credibility. In the multinomial (multi-outcome) case, market-clearing requires solving the eigenvector equation $(P W^\top) m = m$, where $P$ encodes model outcome probabilities and $W$ is hypothetical terminal wealth. Pseudocode for the binary case is as follows: ```python # Initialization B[i] = prior[i] # for i = 1...K, sum B = 1 for t in 1...T: # 1. Read forecasts q = sum(B[i] * p[i] for i in 1...K) # 2. Compute Kelly fractions for i in 1...K: b[i] = (p[i] - q) / (1 - q) # 3. Observe outcome o ∈ {0,1} for i in 1...K: if o == 1: B[i] *= (1 + b[i] * (1/q - 1)) else: B[i] *= (1 - b[i]) # 4. Normalize Z = sum(B) for i in 1...K: B[i] /= Z # At any t, B[i] equals model M_i's posterior credibility ``` [2602.09982] ## Conclusion Kelly betting yields a formal equivalence between wealth maximization via optimal sequential wagering and Bayesian model evaluation. In both theoretical and empirical terms, this approach implements real-time, order-aware, and posterior-consistent updates of model credibility, recovers traditional Bayesian principles in aggregate, and supports discounted or tempered updates via fractional Kelly. Market prices correspond to posterior-predictive means, and the worst-case regret bounds have the interpretation of Bayesian model-selection penalties. Kelly-based Bayesian evaluation thus provides a principled alternative to classical scoring rules for sequential, real-time model assessment and aggregation [1201.6655][2602.09982]. Source: https://www.emergentmind.com/topics/kelly-betting-as-bayesian-model-evaluation