In this 46 pages foundational paper we develop an information-theoretic protocol for systematically quantifying how useful new assets or trading strategies are to investment managers.
Are you an investment manager looking to go beyond traditional linear i.i.d. factor models, or the alpha/beta narrative, to determine whether a new strategy adds value to your existing setup while accounting for non-linearities and temporal dependencies so as to mitigate Black-Swan types of events?
Are you looking for an alternative to Pearson's correlation for quantifying pairwise dependency between time series of returns, one that captures non-linearities and temporal dependencies?
Are you looking for an information-theoretic approach for quantifying the diversification potential in trading new assets that is independent from your allocation process?
Are you interested in an information-theoretic approach for quantifying how predictable a time series is before dedicating any resources to attempting to predict it, and without placing any limitation on which model would work best for predicting said time series?
Are you an investment manager looking for a statistical hypothesis test for whether a new trading strategy performs at least as well as your best deployed strategies, while rigorously accounting for the number of trials to mitigate backtest overfitting?
If you answered yes to any of the above, then you will find in our yellow paper useful machine learning methodological contributions developed with a finance-first mindset.
When it comes to processing massive financial data, machines have a natural advantage over humans.
The human brain did not evolve to process and make sense of vast amount of heterogeneous and sometimes high-frequency financial data. Traditional quant trading models only exploit market inefficiencies (or 'alphas') arising as deviation from simplistic equilibrium models (such as implied by factor analysis) constructed around a simple economic narrative. We believe this approach barely scratches the surface of market inefficiencies.
The more trading becomes electronic, the harder it will be for humans to compete with machines.
The relentless electronification of trading across all asset classes, driven by technological advances, cost reduction, and regulatory pressures, means that virtually every type of market and price-sensitive piece of information will be accessible by computers in real-time in the near-future. This will significantly widen the scope of automation in financial markets. We believe that there is no trading logic or market insight that cannot be learned by an artificial intelligence using the same data as a human being.
We view the automation of trading idea generation as a viable long-term solution to management fee compression.
Investment management and trading idea generation are perhaps the last frontiers of automation in financial markets. As the downward pressure on management fees is here to stay, there is an industry-wide need for more efficiency and scalability in trading idea generation. We believe that artificial intelligence will not only generate investment returns that are expected of active management strategies, but, by utilising the same core technology, artificial intelligence will provide the scalability required to cope with a low fees environment.
We believe that computing has become cheap and powerful enough to enable an AI revolution in finance.
Over the past 50 years, computers have become 10,000 times more powerful per dollar. This development has played a key role in most of the artificial intelligence breakthroughs that have been witnessed over the past decade. Today, specialised hardware such as TPUs and strategic initiatives such as quantum computing, are paving the way for the next artificial intelligence revolution. We believe that the last barrier to an AI revolution in finance is machine learning methodology, and we have set out to break it down.