Online Learning of Portfolio Ensembles with Sector Exposure Regularization

Guy Uziel,Ran El-Yaniv

DOI: https://doi.org/10.48550/arXiv.1604.03266

2016-04-12

Abstract:We consider online learning of ensembles of portfolio selection algorithms and aim to regularize risk by encouraging diversification with respect to a predefined risk-driven grouping of stocks. Our procedure uses online convex optimization to control capital allocation to underlying investment algorithms while encouraging non-sparsity over the given grouping. We prove a logarithmic regret for this procedure with respect to the best-in-hindsight ensemble. We applied the procedure with known mean-reversion portfolio selection algorithms using the standard GICS industry sector grouping. Empirical Experimental results showed an impressive percentage increase of risk-adjusted return (Sharpe ratio).

Machine Learning

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is how to control and reduce risks while improving risk - adjusted returns (such as the Sharpe ratio) in online portfolio selection by introducing a regularization mechanism for industry - sector exposures. Specifically:

1. **Balance between Risk and Return**: Traditional online portfolio selection research mainly focuses on maximizing cumulative wealth or returns, but ignores effective risk management. This paper aims to control risks by encouraging the diversification of portfolios on predefined risk - driven stock groupings (such as industry sectors) while optimizing returns.

2. **Challenges in Dynamic Risk Control**: In an adversarial setting, sub - linear regret based on risk - adjusted metrics (such as the Sharpe ratio) is impossible to achieve. Therefore, this paper proposes a new method to achieve effective risk control under the premise of ensuring a logarithmic - level regret bound.

3. **Application and Verification**: The authors apply this method to the known mean - reversion portfolio selection algorithm and use the standard Global Industry Classification Standard (GICS) for industry grouping. The experimental results show that, compared with directly applying the basic algorithm and previous algorithms that attempt to use group diversification, the new method significantly improves the risk - adjusted returns (Sharpe ratio).

In summary, this paper achieves risk control and return optimization in portfolio selection by introducing a new online learning framework combined with convex optimization techniques, especially achieving remarkable results in terms of diversification at the industry - sector level. This not only improves the theoretical performance of the model but also shows superior performance in practical applications.

### Mathematical Formula Representation

- **Sharpe Ratio**:
\[
S=\frac{R_A - R_f}{\sigma_A}
\]
where \(R_A\) is the annualized rate of return of investment algorithm \(A\), \(\sigma_A\) is its annualized standard deviation, and \(R_f\) is the risk - free annualized interest rate.

- **Loss Function**:
\[
g_t(w)=-\log(\langle X_t, P_t w\rangle)
\]
where \(X_t\) is the market vector in the \(t\)-th round, \(P_t\) is the updated portfolio matrix of the sub - algorithm in the \(t\)-th round, and \(w\) is the allocation vector.

- **Group Norm**:
\[
L_G^{(\infty,1)}(X)=\left\|\left(\|X_1\|_1,\dots,\|X_m\|_1\right)\right\|_\infty
\]
where \(X_i\in\mathbb{R}^n\) is equal to \(X\), but all coordinates not belonging to \(g_i\) are set to zero.

These formulas are used in the paper to describe the key components of the algorithm and its performance analysis.