The USSI and it's relation to Financial Markets and Economic Indicators
Large scale sentiment trackers are often touted as an orthogonal data source (i.e. unrelated to typical economic data) that can be of use for asset managers, banks, and other professionals in the financial markets space. Orthogonal data is something that a lot of people in trading and asset management are looking for these days to gain an edge on competitors who are using traditional data, often in very similar ways. Until recently, it was hard to assess data sources from social media since they covered a relatively small time window, making it hard to test the strength of correlations between measures and financial market data. Now that it’s possible to look at 5+ years of data, it’s possible to take a look to see how our data from the USSI stacks up.
We have been looking closely at two indicators that relate to the mood of consumers and investors, the volatility index (or VIX) and the consumer confidence index (CCI). I’ll let Steve describe how they approached the question, how they conducted their research, and what they found:
I must admit that we started off as skeptics so to keep things manageable we first looked at the methodology underlying the construction of the USSI and focused on estimating sentiment with respect to specific films JANYS Niels discussed in a previous blog post. Our interest was to understand within the film industry, does social media content help forecast business outcomes. There are a number of studies that conducted similar types of analysis and we also consider new findings on model selection and model averaging in high-dimensional problems where the number of unknown parameters is large relative to the sample size. Our main findings are that while social media data can improve box office predictions for Hollywood studios irrespective of the method employed, model uncertainty appears important for this industry. Further, we found that there are benefits from having two distinct measures of social media data, volume and sentiment since they capture different dimensions of product awareness and purchasing intentions. This paper is forthcoming at the Review of Economics and Statistics and an earlier version is available for download from the NBER working paper series.
After completing this paper, we became confident that there was significant potential in using the algorithm to capture something along the lines of mood or intentions. Since this concept was somewhat fuzzy when looking at the aggregate Twitter messages, we next decided to look at the link between the USSI and existing metrics that capture other fuzzy behavioral concepts that have large impacts in finance circles. Specifically, we looked at consumer confidence which is reported monthly and a rapidly changing measure designed to capture expected stock market volatility. Our initial findings is that hourly changes in sentiment measures are strong predictors of the direction in which the VIX moves and that there are strong correlations between the daily USSI and CCI. In fact, our work finds that the correlation is much stronger on the days during which the data underlying the CCI is collected and slightly lower on the day when the CCI is released. These results are suggestive that the USSI will add value to forecasting models for certain financial indices and we speculate that collecting a USSi on market participants will likely add more value than the aggregate USSI.
We are currently completing two papers that formally explore these issues as well as consider new and improved ways to incorporate social media data in forecasting models. In all of work, we have contrasted machine learning to econometric approaches and generally find that hybrid approaches can outperform. As new papers will be completed, we will be sure to update Janys who may post links to the work on the website. In summary, much more work can be done with the USSI and other metrics calculated on subsets of specifics tweets using the methodology underlying the USSI to help develop new tools and strategies to forecast outcomes of interest to policymakers and financial analysts.”
We think Steve and Tian’s work is just the tip of the iceberg in terms of using our data, and we welcome other researchers to look at our data and see what other interesting correlations might be found.