I study the effect of information networks on learning to forecast in an asset pricing market. Financial traders have heterogeneous price expectations (Hommes, 2011), are influenced by friends (Nofsinger, 2005) and seem to be prone to herding (Shiller and Pound1989). However, in laboratory experiments subjects use contrarian strategies (Drehmann et al., 2005; Cipriani and Guarino, 2009). Theoretical literature on learning in networks is scarce and cannot explain this conundrum (Panchenko et al., 2013).

I follow Anufriev et al. (2014) and investigate an agent-based model, in which agents forecast price with a simple general heuristic: adaptive and trend extrapolation expectations, with an additional term of (dis-)trust towards friends’ mood. Agents independently use Genetic Algorithms to optimize the parameters of the heuristic. I consider friendship networks of symmetric (regular lattice, fully connected) and asymmetric architecture (random, rewired, star).

I find the agents to learn contrarian strategies, which amplifies market turn-overs and hence price oscillations. Nevertheless, agents learn similar behavior and their forecasts remain well coordinated. The model therefore offers a natural interpretation for the difference between the experimental stylized facts and market surveys.

Field: econometrics/non-linear dynamics