Bayesian Perspective of Universal Portfolios

Jungbin JunAugust 14, 2026#Online Learning

Let $X \in \mathbb{R}_+^m$ denote a stock market vector, and let $\lambda \in \Delta^m$ denote a portfolio. Given a portfolio $\lambda$, the investor’s one-period wealth is given by

$$ W(\lambda) = \lambda^\top X. $$

Now consider a sequence of stock market vectors $\lbrace X_t \rbrace_{t \in [T]}$. The wealth accumulated by a fixed portfolio $\lambda$ over $T$ periods is

$$ W_T(\lambda) = \prod_{t=1}^T \lambda^\top X_t. $$

The $\mu$-weighted universal portfolio at time $t$ is defined as

$$ \bar{\lambda}_t = \frac{ \int_{\Delta^m} \lambda W_{t-1}(\lambda)\, d\mu(\lambda) }{ \int_{\Delta^m} W_{t-1}(\lambda)\, d\mu(\lambda) }. $$

We may define a distribution $q_t$ over portfolios as

$$ q_t(\lambda) \propto W_{t-1}(\lambda)\,\mu(\lambda). $$

Under this interpretation, $W_{t-1}(\lambda)$ plays the role of a likelihood, while $\mu(\lambda)$ plays the role of a prior distribution.

Indeed, letting

$$ Z_t = \int_{\Delta^m} W_{t-1}(\lambda)\,d\mu(\lambda), $$

we have

$$ \mathbb{E}_{\lambda \sim q_t}[\lambda] = \frac{1}{Z_t} \int_{\Delta^m} \lambda W_{t-1}(\lambda)\,d\mu(\lambda) = \bar{\lambda}_t. $$

Thus, the universal portfolio $\bar{\lambda}_t$ can be interpreted as the posterior mean under $q_t$. From this perspective, universal portfolio selection admits a natural interpretation as a form of online Bayes.

Empirical Bayes

O'Sullivan et al. [1] proposed an adaptive universal portfolio of the form

$$ \tilde{\lambda}_t = \frac{ \int_{\Delta^m} \lambda \alpha_t(\lambda) W_{t-1}(\lambda)\,d\lambda }{ \int_{\Delta^m} \alpha_t(\lambda) W_{t-1}(\lambda)\,d\lambda }, $$

where $\alpha_t : (\Delta^m, \lbrace X_j \rbrace_{j \in [t-1]}) \to (0,1)$ is chosen to be a non-decreasing function of the past wealth $W_{t-1}(\lambda)$.

The factor $\alpha_t(\lambda)$ can be viewed as inducing a data-dependent prior over portfolios. Consequently, the adaptive universal portfolio $\tilde{\lambda}_t$ admits an empirical Bayes interpretation: the prior distribution is adapted using the observed market history, and the resulting portfolio is the posterior mean under this data-dependent prior.

I am not yet sure how useful this perspective will be, but I believe it may lead to interesting insights.

References

[1] O'Sullivan, Patrick, and David Edelman. "Adaptive universal portfolios." The European Journal of Finance 21.4 (2015): 337–351.

Back to Posts