Publication
This research project is an experimental study of decision-making in very difficult contexts resembling those encountered in financial markets. The starting point was the empirical observation that financial assets are objects of a very complex kind. Specifically, financial assets generate rewards in an unpredictable way (stochastic processes), and they jump regularly (unstable processes). What's more, they are covert: finance practitioners lack information about their nature, and may well be unable to uncover it. After all, can humans really learn the nature of unstable stochastic processes? To address this question, I designed an investment task, the "Boardgame," characterized by assets – in the form of locations displayed on a board – which are, like the financial assets in the real world, unpredictable, unstable, and (initially) unknown. Specifically, the game consists of choosing between six locations of initially unknown probability of winning or losing. Each trial, the player selects one location and immediately receives the outcome returned by the chosen location (a reward, a loss, or 0 CHF). She accumulates gains and losses throughout the game, with the goal of maximizing the cumulative earnings. One essential characteristic of the Boardgame is that outcome probabilities – at all locations – jump regularly. An ideal player accounts for jump occurrence and re-learns outcome probabilities once a jump has occurred. Such two-stage mental process – consisting of jump detection and subsequent reappraisal of the probabilities – is extremely demanding. I invited 62 subjects – all undergraduate or graduate students at the EPFL – to play the Boardgame during 30 minutes, and recorded their action in each trial (namely, which location they chose). From their choices, I attempted to infer the nature of their learning in the game. To do so, I followed an inter-disciplinary approach, at the intersection of economics, machine learning, and neurobiology. I first formalized normative learning in the game (how ideal players learn), as well as bounded rational learning (how adaptive – intelligent – agents, albeit of limited cognitive capabilities, are expected to learn). I then examined whether my subjects did a good job of learning in the Boardgame, by comparing the fits of the normative and bounded rational models. I eventually studied the neural implementation of the behavior at work. It appears that for the majority of my subjects, the normative learning model did a much better job of explaining actual behavior than the alternative, bounded rational model. In other words, subjects were apt at learning the outcome probabilities of the locations, despite the difficulty of the enterprise. This suggests that people may be better fit to cope with the instability of financial markets than previously thought. Beyond being of interest for the finance field, this result has actually a broader scope. For it demonstrates the human ability to adapt behavior to suit