ExtractAlpha, a quantitative research firm seeking to extract signals from alternative data, recently launched a new product based on consumer behavior mined from product reviews, search queries and social networks.
ExtractAlpha is comprised of ex-StarMine employees. Founder Vinesh Jha was head of research for StarMine from 1999 to 2005, when he joined the proprietary trading desks at Merrill Lynch and Morgan Stanley. Before founding ExtractAlpha he was with PDT Partners, a spinoff of Morgan Stanley’s quant prop trading group.
ExtractAlpha is focused on building models on alternative data: “We are interested in any quality data set with sufficient historical and cross-sectional coverage which we hypothesize should intuitively have predictive power in forecasting equity prices,” said Jha in a recent interview.
The firm recently released a new model based on a stock selection score derived from various data types across web site, search and social platforms. The data is collected by a new firm, alpha-DNA. On a monthly basis, alpha-DNA tracks more than 75 billion digital consumer interactions to identify changes in consumer demand for products from 2000 publicly traded companies. Historical data is available back to 2011.
ExtractAlpha says top ranked stocks in the model have beaten revenue expectations 71% of the time between 2012 and 2015.
The firm offers two other models: one based on TipRanks’ proprietary database of analyst and blogger recommendations (the model focuses only on blogger recommendations since firms like StarMine already measure analyst recommendations); and a short-term trading model based on the reversal effect along with other factors.
Extract Alpha differs from other alternative data providers that are focused on mining data sets to complement fundamental analysis. Pioneering alternative data firm Majestic Research initially focused on marketing data to quant firms but shifted tactics by hiring analysts to interpret the data.
ExtractData is solely quant-oriented, with the philosophy of saving quant investors the time and effort of developing their own models and, more importantly, eliminating the frustration of licensing data sets only to find that they do not have sufficient signal to be useful inputs.