Using machine learning to segment investor behaviour

Purpose Investors saving for retirement have an unfortunate tendency to switch investments (more so in volatile periods), which leads to worse ex-post investment return outcomes – the “behaviour tax”. This paper uses machine learning clustering algorithms to better understand this behaviour. Design/methodology/approach It improves the clustering efficacy of a previous study by Nixon and Gilbert (2022) in several ways: by using a different clustering algorithm (k-means), adding a new variable (portfolio drawdowns) to the feature set, using a large new dataset of investor switching behaviour which includes a period of extreme investment volatility (2020) and clustering on six separate calendar years (instead of only a pooled sample). Findings The seven behavioural clusters identified are more clearly defined with different levels of behaviour tax or value eroded by switching between mutual funds. The algorithm organically differentiates the population into different groupings of investors, each incurring different levels of behaviour tax. Research limitations/implications (1) We don’t know whether the switch was initiated by the adviser or the client. (2) We don’t know in all cases whether the switch is behavioural in nature (i.e. not as a result of a changing goal). There are no causal relationships between the behaviour and feature set, only correlational. Practical implications Financial services firms using this methodology can more effectively segment their customer bases, allowing for more personalised engagement. Originality/value This research deals with a large customer database, which lends itself to the application of machine learning techniques to better understand customer behaviour. In doing so, more effective and personalised engagement is possible and manage/reduce/eliminate the behaviour tax.