Predicting Stock Market Using Online Communities Raw Web Traffic Streams

This paper investigates the predictive power of online communities traffic in regard to stock prices. Using the largest dataset to date, spanning 8 years and almost the complete set of SP500 stocks, we analyze the predictive power of raw unstructured traffic by filtering stock daily returns with traffic features. Our results partially challenge the assumption that raw traffic simply trails stock prices, as expected from a noisy signal without the sentiment direction. Raw traffic is shown to predict prices with statistical significance but with small economic impact. Anyway, this impact rises to moderate under the following conditions: 3 to 7 days lag and stable traffic level. Moreover, the quality of the predictions significantly increases when a high level of traffic is coupled to low market volatility, while a high level of traffic in period of high volatility usually denotes late reactions to violent market movements and a consequent poor predictive power. The findings set interesting future works in the definition of novel indicators for market analysis based on web traffic analysis, to be coupled with complementary tools such as sentiment analysis.