Computers can be clever, yes, but surely they lack any inherent sense of self, humanity or ability to act fiduciary, at creation? Could they? Hot on the heels of robo-advice, robo-selection, as it relates to the selection and research of mutual funds, is the next frontier for FinTech. As more assets move to index-based and systematic strategies, so the dynamics of fund research are changing. How then can robo-selectors put the client above all others? Friedrich Nietzsche wrote that morality was the herding instinct of the individual . If true, then that herding behaviour could be defined and mapped.

The concept of “fiduciary” certainly goes well beyond selecting the cheapest passive fund and leaving it there for 25 years. Fiduciary has also expanded into political, society and environmental levels through industry initiatives like the UN Principles for Responsible Investment (PRI) and Sustainable Development Goals. However, the very discharge of fiduciary itself could be more transparent, for example does an investor require a human to offer a fiduciary service or indeed to be a fiduciary fund selector? Could computers act in a fiduciary manner? First, consider that human decisions litter history with malfeasance: Madoff, Arch Cru, boiler rooms, LIBOR-rigging, flash boys and so on.

Computers are not predisposed to wrongdoing, only programmed to do so – a thought for the rapid proliferation of robo. With the rise of the robots, the battle for white-collar roles has begun. White-collar roles by and large have little physicality to overcome, lots of talking, thinking, writing reports and drinking coffee. Martin Ford’s book¹ aptly summarizes the bleak outlook for white-collar finance professionals; a second digital age if you will. The asset management industry (“the City”) is actually one of the last bastions of human industrial intensity, be it actuaries, fund managers, brokers, consultants or professional fund investors, and the myriad of pseudo-executives and support staff in between.

What does being a fiduciary require? Empathy, diligence, determination, trustworthiness, expertise, judgement. Putting the investor before yourself. These may seem all too human and a realm beyond digitization yet, but when Google’s DeepMind AlphaGo program beat South Korea’s Lee Sedol 4–1 in the first of a series of games in Seoul in March 2016, the fiduciary opportunities from adaptive artificial intelligence (AI) took a huge step forward.

Over 2500 years old, in terms of rules, Go has considerably simpler rules than chess. Black and White sides each have access to black and white stones on a 19 × 19 grid. Once placed, stones don’t move. The aim of the game is to completely surround and capture opposite stones, which are then removed from the board. Here the complexity and artistry arises – organic battles from the corners to the centre of the board. Given that a typical chess game has a branching factor of about 35 and lasts 80 moves, the number of possible moves is vast, about 35⁸⁰ (35 to the power of 80) possible moves, aka the “Shannon number”. Go is much much bigger. With its breadth of 250 possible moves each turn, and 150 likely moves, there are about 250¹⁵⁰ (250 to the power of 150) possible moves.

In many ways the computations of a fund selector are far more complex than a Go stratagem, but the small judgemental biases are similar. Asset allocation, over- and underweighting assets against a “neutral”, giving rise to tactical positions, has its very roots in war strategy and hence Go. Of course, fund selection is neither science nor art, it combines fuzzy logic with data inputs into a decision.

Imagine if fund selection can be derived from software: a neural network computing program that can screen thousands of funds and make judgements, shortlist recommendations, assess suitability and compatibility against a mandate or investor needs and monitor the outcome of those decisions, all the while underwriting the capability of a fund manager based on a model, new information and past data, recommending changes while simultaneously weighing the impact on the portfolio and topical metrics like turnover, cost, etc. Performance analysis has already been digitized; other types of information can be stored and used for learning, algorithms can replicate judgemental nudges and biases based on common material changes like manager experience, tenure, benchmark, fund changes, moving firm, news flow and so on. In many ways software can far better interpret the signals, and its significance needs to lay a good starting baseline for self-learning. In many ways the components already exist as separate FinTechs, simply needing a complex adaptive algorithm to link it all together.

The other challenge is the retention of experience, digital wisdom if you will. Here AlphaGo solved through coupling self-learning with a differentiable neural computer (DNC). Memory requires memory, literally. The DNC architecture differs from others by selectively writing memory as well as reading, allowing iterative modification of memory, according to its developers. All mind-bending stuff for a mere fund analyst!

Firstly, robots follow three guiding laws: the three Laws of Robotics (“Asimov’s Laws”) devised by the science fiction author Isaac Asimov. These have come to be fundamental guiding principles for the robotics industry and so too should apply to the fiduciary robo-advisor and robo-selector:

• A robot may not injure a human being or, through inaction, allow a human being to come to harm.
• A robot must obey the orders given it by human beings, except where such orders would conflict with the first law.
• A robot must protect its own existence as long as such protection does not conflict with the first or second laws.

The fiduciary duty thus falls to the programmer of the algorithm, who instructs the programme to make decisions. Taking these laws, it is not unfathomable that computers can be programmed and learn to put the interests of the client first and foremost.

1. Uphold a fiduciary standard; all conflicts of interest must be disclosed. A computer has no conflicts unless they are first programmed. Like a driverless car, its function is to serve the purpose without question.
2. A fiduciary has a “duty to care” and must continually monitor not only a client’s investments, but also their changing financial situation. A computer can monitor 24/7 continuously and is not restricted by fatigue or the adviser’s/fund selector’s diary. A sequence can be included if the client does not supply an update within x days or could be linked to the client’s accounts, email, diary and so on.
3. Understand changes to a client’s risk tolerance, perhaps after a painful bear market. Perhaps there was a family change. Under the suitability standard, the financial planning process could begin and end in a single meeting. For fiduciaries, that first client meeting marks only the beginning of the legal obligation. We have seen the term “orphan clients”, and humans have a great track record of dropping less profitable clients (value pools).
4. Monitor, learn, adapt and assess fund changes. The reality is that many fund investors do not monitor their decisions often enough or with objectivity. They are susceptible to heuristic biases. Yet a computer can continuously monitor cost, turn- over, risk, changes and performance. It can monitor RSS feeds, performance, fund manager commentary, portfolio positions, information supplied by the client, instructions, deal flow, thousands if not millions of data points analysed through neural networks.

From the above, the algorithm can include a core hierarchy of neural paths and decision trees: efficacy of active management versus benchmark/passive based on a learning empirical algorithm and relative valuation of index momentum, the carbon impact of a portfolio, the environmental, social and governance score of the fund and underlying stocks, Stewardship score, economic value comparison of available universes and fund structures, risks relating to the firm – like fines or corporate restructures, rating changes, flow data, people changes, process issues, significant portfolio turnover or positioning changes, risks, performance issues, ongoing costs. From these broad sections, the algorithm can create thousands of rules, rankings, logic maps and decision points driving further questions or a BUY, HOLD or SELL decision.

The move towards more fiduciary transparency is an important step towards digitizing robo-selection. The danger is that self-learning AI itself compromises the fiduciary obligation to the investor and it is here that Asimov’s laws need to be hardwired, tested and continuously measured. Perhaps the answer is more symbiotic than we might be led to believe. A third way? If you feel the notion of digital fiduciary is still more science fiction than threat, then you may be right but we do well at becoming better “cyborgs”² – using technology, using the cloud, the crowd, to generate positive fiduciary evidence for the value added by human fund selection. Robo is Go, your move!