Finding Alphas, 2nd Edition

12Techniques for Improving the Robustness of Alphas

By Michael Kozlov

INTRODUCTION

The main goal of alpha research is to predict and try to outperform the market. However, most investors desire not only returns but also low risk exposure and some confidence in their ability to anticipate trading results. With these requirements in mind, we have defined a robust alpha.

A robust alpha should have the following properties:

• Invariance under modification of the traded universe: An alpha should reflect a statistical property that is independent of the choice of specific instruments to trade. Changes to the instrument set are frequently imposed by the market as a result of regulatory changes, customer preferences, liquidity decreases, short bans, etc.
• Robustness to extreme market conditions: An alpha should not have excessively sharp declines in any of its performance benchmarks. The most common benchmarks used for alpha performance measurement include information ratio, maximum drawdown, and return.

METHODS FOR ROBUSTNESS IMPROVEMENT

In this section, we will introduce the most common and established techniques for improving the robustness of an alpha. In general, our goal is to ensure stable alpha performance that is not unduly affected by input data outliers or other small departures from model assumptions.

Methodologies for robustness improvement can be classified into three categories: ordering methods, approximation to normal distribution, and limiting methods. Below we will ...