4. Correlation Analysis
A summary of correlation tests between VIX and Metaculus Data
Overview
Our next step was to study the strength and extent of correlation between our respective timeseries. To do so, we looked into a handful of statistical methods ranging from pearson correlation and granger causality to autocorrelation and dynamic time wrapping methods between the returns of each dataset. A complete write up of our correlation methodology and code can be found at Correlation Guide and return_corr.py script respectively.
Our strongest correlations appeared to occur at timelags approximately 10 months in advance of the financial timeseries. While some of these correlations were quite strong, due to the size of the lag, we cannot resonably conclude a causal realtionship. Instead, we hypothesize that the Metaculus user data only predicts spikes in market volatility during times of great unrest and is largely uncorrelated during all other times. As such, on the whole, the metaculus timeseries does not provide much predictive power on market volatility, but we still may be able to isolate specific periods of strong correlation.
