Assessment of learning to forecast experiments in multiple-asset markets

Abstract

We conducted six Learning-to-Forecast experiments as an exploratory case study on expectation formation and behavioural dynamics in multi-asset markets. Participants acted as financial advisors and forecasted the prices of three risky assets, using initial price histories as their primary source of information. One asset exhibited a stable initial price history, while the other two were more volatile. The aim of the thesis is to investigate how initial price history influences overall market dynamics, coordination of expectations, and individual forecasting strategies in multi-asset experimental markets. This design allows us to move beyond the traditional single-asset focus and explore complex inter-asset interactions. Regarding overall market dynamics, we observed two distinct patterns: stable markets, where prices remained close to fundamentals with low volatility, and markets with moderately large bubbles. Initial price history played a key role in shaping market prices – the asset with a stable initial price history consistently showed lower volatility. Moreover, a negative dependence between two assets during the initial price history persisted mostly throughout the experiments and influenced overall market behaviour. Generally, participants were highly coordinated. Alignment of expectations tended to strengthen during price increases but weakened sharply after sudden declines. This likely reflects diverging beliefs about future price movements and was most evident in markets with moderately large bubbles. The asset with a stable initial history generally showed higher coordination among participants. We also conducted a comprehensive analysis of individual forecasting strategies. First, our findings reveal a clear dominance of the trend-following heuristic. Second, the majority of participants consistently applied the same heuristic across all assets. Third, panel regression analysis confirmed that participants primarily based their forecasts on the historical behaviour of the target asset. Additionally, small but statistically significant cross-asset influences were detected, indicating that participants incorporated price information from other assets when forming their forecasts.