Does web anticipate stocks? Analysis for a subset of systemically important banks NTTS March 11, 2015 www.jrc.ec.europa.eu
Rationale ‘ … people sometimes trade on noise as if it were information’ (Black, 1986) Where can we find the “noise” In the web 1. Is web buzz able to lead stock behavior (for banks)? 2. Are stock movements sensitive to the geo-tagging of the web buzz
Steps of the experiment 1. Scrape the web 2. Look for texts containing given keywords 3. Extract the mood of each text (sentiment analysis) 4. Analyse the relationship between the mood and stock movements
Scrape the web JRC- Europe Media Monitor Born in 2002 Scrapes more than 10,000 RSS feeds/HTML pages from 4000 media websites worldwide (EU+US sources are 1500 - from 140 for Germany, to 50 for Spain) retrieves about 200.000 new news articles per day Keywords based (over 1500 categories) Works in 60 languages (sentiment analysis in 14 languages) Real time: scrapes the net every 10 minutes , 24h/day Attracts up to 1,2 Million hits per day http://emm.newsbrief.eu
Added value Novelty with respect to existing literature 1. First to analyse banks (objective: anticipate turbulences) 2. Multilingual sentiment (usually literature on English web texts) 3. Full control of sources (usually literature works with a limited number of texts from a source) 4. Geo-tagging: first to analyse which information matters More on Nardo et al. 2015 Journal of Economic Surveys
Analysis Web and stock daily data: from Dec. 2013 to April 2014 (overall 100 days of trade) Subset of 10 banks (Barclays, BBVA, BNP Paribas, Crédit Agricole, Deutsche Bank, HSBC, Royal Bank of Scotland, Santander, Société Générale and Unicredit) Each combination of 8 stock prices variables, 12 web buzz variables, 4 set of sources (with different geo-tagging), various stock markets. The relationship between stock data and web news is analysed via cross-correlation function, • Granger causality (rank-sum test) • Factor and Cluster analysis •
Cross correlation Results Average (10 banks): between 0.33 and 0.37 at lag δ=0
Results Average correlation between number of web-texts and various stock variables EU stock exchanges NYSE Location of information matters
Effects of geo-tags For each bank we estimate the equation: 𝑇 𝑢 = 𝛽 + 𝛾 1 𝑇 𝑢− 1 + 𝛾 2 𝑋 𝑢 + 𝛾 3 𝑋 𝑢− 1 + 𝜁 𝑢 Estimation for 4 sources of web buzz (W): 1. Sources located in EU+US 2. Sources located in EU 3. Sources all over the world (ALL) 4. Sources located in the country where the banks has headquarters (country) For each estimated model we calculate the % change in the model fit (R2) using option 4 as baseline
Results sources EU-US vs All vs EU vs Country Country Country Barclays 24.1% 23.0% 21.8% BBVA 21.2% 5.2% 17.1% BNP Paribas 24.6% 29.8% 23.9% Crédit Agricole 3.3% 1.3% 2.5% Deutsche Bank 22.4% 24.7% 22.0% HSBC -26.5% -26.9% -28.9% Royal B. Scotland 27.5% 29.7% 29.5% Santander 4.8% -0.2% 3.5% Société Générale 11.1% 2.4% 6.1% Unicredito 14.1% 14.9% 11.6%
key where web gain wrt variable is the Results Results anticipates uniformed anticipa key stocks investor ted informa Barclays On average web does BBVA 4% prices EU+USA not Granger cause Stock BNP Paribas Crédit Agricole 4-9% volatility EU+USA But we find good prices Results for individual Deutsche Bank 5-7.5% and EU+USA banks volumes HSCB yal Bank of Scotland 5% prices EU+USA Santander Société Générale 5% prices EU+USA prices Unicredito 4-11% and EU+USA volumes
The project is on going future planning: • Analyse the weekly averages (to eliminate some noise) for 25 banks • Investigate general trends in the Euro area (via alert setting) • More ambitious: text mining on keywords michela.nardo@jrc.ec.europa.eu
Recommend
More recommend
