Applying Minimum Message Length to the Clustering of Mutual Funds

Clustering has been widely studied to group data into clusters. Several methods have been used including Maximum Likelihood (ML), Information Criterion by Akaike (AIC), and Bayesian Information Criterion (BIC) by Schwarz. In this paper, Minimum Message Length (MML) is applied to the clustering of mutual funds data. In this application, data are assumed to come from multivariate correlated Gaussian distribution. For this, MML principle needs to be numerically approximated. The modeling results are contrasted with those obtained using alternative methods, in terms of probability-bit costings and clustering structures. The experiment's findings demonstrate that, in terms of the fitted probability bit costings, MML clustering provided a more trustworthy model than AIC and BIC with significantly less bits required in conveying the data given the model. MML clustering also handled overlapping clusters better compared to modeling using the combination of ML with AIC and BIC. Furthermore, mutual funds trading have shown changes of movements during the pandemic Covid-19 with performances of mutual funds tend to be decreasing across funds, especially during the first 15 months of the period. Only several funds were grouped differently compared to most funds analyzed. The latter have shown effect of pandemic Covid-19 the most with lower returns compared to the returns of most funds.