Getting More for Less – Better A/B Testing via Causal Regularization

Getting More for Less – Better A/B Testing via Causal Regularization

Used in management and engineering, an Ishikawa diagram shows the factors that cause the effect. Smaller arrows connect the sub-causes to major causes.

Causal regularization solves several practical problems in live trading applications:

estimating price impact when alpha is un- known and estimating alpha when price impact is unknown. In addition, causal regularization increases the value of small A/B tests: one draws more robust conclusions from smaller live trading experiments than traditional econometric methods. Requiring less A/B test data, trading teams can run more live trading experiments and improve the performance of more trading algorithms.

Using a realistic order simulator, we quantify these benefits for a canonical A/B trading experiment.

The interplay of information and trading is critical to trade execution. Practitioners refer to price moves caused by their trading as price impact and price moves independent of their trading as alpha.

An essential corollary is that trading causes price moves that otherwise would not have happened. Moreover, successful investment strategies across all asset classes trade to minimize the price impact and maximize the alpha during trading.

The literature highlights a crucial challenge when estimating impact: alpha signals cause trades!

“The larger the volume Q of a metaorder, the more likely it is to originate from a stronger prediction signal.”

The industry standard for addressing this bias is through controlled live trading experiments, such as A/B tests that randomize decisions.

A/B tests address trading biases but present three downsides:

Firstly, one discards the bulk of their trading data.

Secondly, there are far fewer trades without alpha than with alpha, making it challenging to estimate high dimensional models using machine learning.

Finally, the submission of alpha-less trades leads to additional trading costs.

The research of Westray & Webster uses causal regularization, to address these shortcomings:

Read The Full Paper: Getting More for Less – Better A/B Testing via Causal Regularization by Kevin Webster, Nicholas Westray :: SSRN

Back To Math & Technology

KEVIN WEBSTER
Imperial College London
NICHOLAS WESTRAY
Head of Execution Research, AllianceBernstein Multi-Asset Solutions &
Financial Machine Learning Researcher, Courant Institute of
Mathematical Sciences, NYU

Getting More for Less – Better A/B Testing via Causal Regularization