Technology

AGNOSTIC: building a bridge between past and future.

Let’s first recapitulate the hurdles encountered when seeking to build a model to support future decisions, as described in the chapters covering opportunity and history.

  1. Almost invariably backtesting ends up overfitting data, even if you are aware of it and take mitigation measures.
  2. An inability to exploit even the smallest portion of the realized performance observed in a backtesting.
  3. Backtesting only tells you what could best have been done in the past and says nothing about the future. Accepting validity of the results for the future is an act of faith.
  4. Selecting the right model to describe the behaviour of potential investments is an arduous task.

To overcome these hurdles, the proposal is to use a sequential allocation decision tool, named AGNOSTIC, taking the most basic and widely available information, returns, as its only input. We no longer want to invest guided by an overfitting of data. Preferring to invest progressively, starting small early and scaling up as confidence builds, rather than investing fully, but too late.

Since confronting the future involves a leap of faith, you may think of AGNOSTIC as a tool that will carry out a reality check at each time step. Starting from day one, updating its trust in each available choice by encompassing relevant history. In doing so, AGNOSTIC will generate a portfolio whose updated weights will help to match, as closely as possible, the performance of the best choice(s), only to be known with the benefit of hindsight.

This is nothing else than an extension of the concept of Universal Portfolios we mentioned earlier. To make a long story shorter, the idea is to weight the investments according to their cumulated performance. As you may know the sum of a set of exponentially growing quantities tends towards the quantity with the largest growth rate.1

AGNOSTIC is precious support for human decision taking, which more often than not happens in the context of a group. An Investment Committee for example, where complex decisions are usually taken by members through majority vote. It is worth remembering that when we are presented with many choices we first feel paralysis and struggle to muster and objectively rank all relevant elements to reach a decision.23 Then, there is a high likelihood we will feel regret, as in hindsight the choice might have been better. Devoid of such human traits, AGNOSTIC claims to boost the investment process. AGNOSTIC is an ensemble-learning algorithm as it relies on a set of underlying learning algorithms, combining them to ensure diversity of learning and decision behaviors. AGNOSTIC also is a meta-learning algorithm, which means it is geared towards learning how to learn.

AGNOSTIC does not rely on any assumption or choice of parameter. The algorithm is the exact same for all clients, whatever the data: asset classes, trading strategies, hedge funds, etc. It does not assume any particular behaviour for the investment, e.g. momentum or mean-reversion.

To cover the needs of finance practitioners, AGNOSTIC can handle missing data, investments with a deferred starting date or different ending date, i.e. investments with different history lengths are handled in a consistent way.

AGNOSTIC is available as Software as a Service (SaaS) with access through a web interface for manual handling or through an API for automated processes. Encrypted communications with the server ensure total confidentiality of your data. A plug-in to handle portfolio constraints is available as an option.

  1. \(\lim_{T->\infty} \frac{\log(\sum_i \exp(\mu_i T))}{T} = \max_i \mu_i\)↩︎

  2. The paradox of choice or why more is less↩︎

  3. Iyengar, Sheena. How Much Choice is Too Much? Contributions to 401(k) Retirement Plans. In Pension Design and Structure: New Lessons from Behavioral Finance. Ed. O. S. Mitchell and S. P. Utkus. New York: Oxford University Press, 2004.↩︎