Online Search Queries and Investor Sentiment: Financial Applications

Lorenzo Tonelli

Student thesis: Master thesis


This thesis examines the use of online search queries as a proxy for investor sentiment and it evaluates their ability to forecast (I) trading activity, (II) abnormal stock returns and (III) implied volatility. Prior research has highlighted that online search query data can be used to measure the attention of unsophisticated investors. Based on this insight, it is reasonable to expect search queries for traded companies to carry information which can predict financial market dynamics. The sample chosen consists of S&P500 constituents on a period ranging from 2007 to 2015. For each firm in this sample, search query data is obtained from Google in the form of Search Volume Index (SVI). Then, three studies are conducted in order to assess SVI’s forecast capabilities. In study (I), the relation between SVI and trading activity is measured by computing a set of time-lagged cross correlation coefficients. In addition, a series of Granger-causality tests is conducted in order to ensure the robustness of the results. In study (II), SVI’s capability to forecast abnormal returns is evaluated by simulating a series of long-short trading strategies based on SVI observations. Abnormal returns of each strategy are then computed by correcting for the most commonly recognized risk factors. In study (III), it is tested whether SVI can improve implied volatility forecasts: several implied volatility autoregressive models AR(p) are estimated in order to provide benchmark measurements; then, these models are augmented with SVI information. The forecasts produced by benchmarks and augmented models are compared and their accuracy is assessed with a series of indicators such as Mean Squared Prediction Error and Mean Absolute Percent Error, among others. The three studies conducted indicate that (I) SVI anticipates and Granger-causes trading activity, (II) SVI incorporates information that translates in abnormal stock returns, but exploiting this phenomena is very difficult because financial markets quickly absorb such information and react accordingly. Lastly, (III) for given specifications of AR(p) models, SVI can improve implied volatility forecast both in-sample and outof-sample.

EducationsMSc in Finance and Strategic Management, (Graduate Programme) Final Thesis
Publication date2016
Number of pages71