The relation between the trading activity of financial agents and stock price dynamics

Transcription

1 The relation between the trading activity of financial agents and stock price dynamics Fabrizio Lillo Scuola Normale Superiore di Pisa, University of Palermo (Italy) and Santa Fe Institute (USA) Econophysics and Networks across Scales -Leiden, May 27, / 15

2 Motivations Financial markets are complex systems where a large number of agents interact through trading A series of stylized facts of the aggregate market quantities (price, volatility, volume, etc.) has been identified and modeled with statistical/time series approaches and with agent based models These facts are the results of the interaction of agents and the effect of external factors (e.g. news), but the connection micro-macro is relatively less explored due to the lack of data We have the thermodynamics (aggregate laws), we lack the mechanics (microscopic laws) and the statistical physics (relation to aggregate), at least from the empirical point of view. 2 / 15

3 A (biased and very personal) selection of previous research with brokerage data Several papers analyzing brokerage data: The ecology of brokers organizes around two classes following in opposite way returns (Lillo et al. 28) Fat tailed distribution of the size of metaorders (Vaglica et al. 28, Toth et al 21, Vaglica et al 21) Market impact of large metaorders is described by a square root function of volume and a 2/3 decay after the completion of the order (Moro et al. 29) Market impact of individual trades is counterbalanced by the price response of the rest of the market (Toth et al. 212) Differential trading behavior of brokers in the electronic and in the over the counter market (Carollo et al. 212, see talk of Mantegna) The long memory of the sign of market orders is due to order splitting rather than to herding (Toth et al, preprint 211). 3 / 15

4 This talk I will present the results of two papers where we investigated empirically databases containing the trading behavior of individual agents Herding impact: Number of agents and stylized facts in financial markets (with G. Germano, V. O. Arbuzov, and S.V. Ivliev, in preparation) How news affect the trading behavior of different categories of investors in a financial market (with S. Micciché, M. Tumminello, J. Piilo, and R.N. Mantegna, arxiv:127.33) Another paper (not discussed here) uses agent resolved data to identify clusters of agents behaving in similar ways (statistically validated networks) Identification of clusters of investors from their real trading activity in a financial market (with M. Tumminello, J. Piilo and R. N. Mantegna, New Journal of Physics 14 (212) 1341). 4 / 15

5 Herding impact: Number of agents and stylized facts in financial markets Research questions How does the number of trading agents scale with the considered time interval? What is the distribution of agents time scales in the financial market? What is the relation between the imbalance between number of buyers and sellers and price return? What is the relation between number of trading agents and volatility? How do the answer depend on the time scale? What is the time lagged relation between these quantities? Since we do not have yet published the paper, in the following slides the results of the first paper (Germano et al.) will not be presented 5 / 15

6 How news affect the trading behavior of different categories of investors in a financial market Research questions What is the the relative role of endogenous and exogenous factors affecting trading behavior of agents? Endogenous factors: price returns and volatility. Exogenous factors: number of news and sentiment via semantic analysis of news. Is there a difference in the importance of these factors between different categories of investors (e.g. households, companies, governmental, or financial institutions)? 6 / 15

7 The Finnish database Central register of shareholdings for Finnish stocks and financial assets in the Finnish Central Securities Depository. Six main categories: non-financial corporations, financial and insurance corporations, general governmental organizations, non-profit institutions, households, and foreign organizations. Foreign investors can choose to use nominee registration, giving aggregate results. Our focus is mainly on Finnish investors. We consider the stock Nokia in the period Jan. 2, 23 - Dec. 3, 28 (1, 51 trading days). The time resolution is one day. Table: Summary of the number of investors (# ids), the number of transactions (N), and the exchanged volume (V ). Volume is given in millions of shares. Category # ids N V Companies 8,396 1,9,226 4,825 Financial 392 4,79,174 21,42 Governamental ,278 1,985 Non profit , Households 129,952 1,555,96 1,993 Foreign 1,45 789,552 7,685 Total 141,19 7,494,14 38,138 7 / 15

8 The Thomson Reuters database and the sentiment proxy Headlines of the NewsScope archive of news released in English by Thomson Reuters. We extract all headlines in English language labeled with at least one Nokia Reuters Instrument Code 11, 484 unique headlines. We consider only the headlines during European trading hours (from 8. am to 4.3 pm UTC time). We construct a sentiment proxy using the number of positive and negative words present in each headline. Positive and negative words are detected by using the General Inquirer from the Harvard psychosocial dictionary. news rate (min -1 ) Opening European markets Opening NYSE Closing European markets Closing NYSE Hour (UTC) Figure: Average daily pattern of the arrival rate of news on the Nokia company. The rate is measured in number of headlines per minute. 8 / 15

9 Investigated variables Investor variables For each day we classify each agent in buyer (B), seller (S), or buyselling (BS) by using the q(i, t) function defined above. N K B (t), NK S (t), and NK BS (t) are the number of investors of category K classified at day t as buyers, sellers or buysellers, respectively. From these variables we obtain N K (t) = N K B (t) + NK S (t) + NK BS (t) N K A (t) = NK B (t) NK S (t) N K R (t) = NK B (t) NK S (t) N K (t) Endogenous variables Daily return Daily volatility (range) Exogenous variables Number H(t) of Nokia headlines Absolute and relative sentiment of the news in a given day S A (t) = G(t) B(t) number of investors of category K excess of buyers of category K relative excess of buyers of category K G(t) B(t) S R (t) = G(t) + B(t) where we use the number of positive (G(t)) and negative (B(t)) words in the headlines. 9 / 15

10 News, volatility, and agents activity H Vol N K N K Financial Households Trading day Figure: From top to bottom the figure shows the time series of the number of Nokia headlines H(t), the daily volatility Vol(t) of Nokia stock, and the time series of N K (t) for the category of Financial investors and for the category of Households investors. 1 / 15

11 Sentiment, return, and buyers excess S R Ret N K R N K R Financial Households Trading day Figure: From top to bottom the figure shows the time series of the relative sentiment indicator S R (t), the daily return Ret(t) of Nokia stock, and the time series of N K R (t) for the category of Financial investors and of Households investors. 11 / 15

12 Regression results and partial correlation analysis: news and volatility Number of news is correlated with volatility, Corr[H, Vol] =.51. We fit N K (t) = α H Ĥ(t) + α Vol Vol(t) + ɛ(t) where x is the standardized versions with zero mean and unit variance of x. Table: Summary of the results of the linear regression of the number N K of trading investors versus the news intensity signal H and the volatility proxy Vol. The number in parentheses are the 5%-95% confidence intervals under Gaussian hypothesis and by using bootstrap analysis. The last two columns show the results of the partial correlation analysis. Investor α H α Vol % variance ρ(n K, H Vol) ρ(n K, Vol H) category of residual of N K Companies.271 (.229,.313).517 (.475,.559) 51.8 % bootstrap (.25,.335) (.437,.597) Financial.195 (.149,.242).479 (.433,.526) 63.8 % bootstrap (.125,.264) (.47,.558) Governmental.238 (.183,.292).192 (.138,.246) 86. % bootstrap (.164,.33) (.119,.262) Non profit.319 (.269,.369).27 (.22,.32) 73.9 % bootstrap (.249,.394) (.199,.344) Households.226 (.188,.263).627 (.589,.664) 41.4 % bootstrap (.165,.285) (.554,.697) Foreign org..158 (.19,.27).442 (.393,.492) 7.9 % bootstrap (.94,.224) (.374,.517) 12 / 15

13 Regression results and partial correlation analysis: sentiment and returns Sentiment is correlated with returns, Corr[S A, Ret] =.155 and Corr[S R, Ret] =.118 (statistically significant). We fit N K R (t) = α S R Ŝ R (t) + α Ret Ret(t) + ɛ(t) where x is the standardized versions with zero mean and unit variance of x. Table: Summary of the results of the linear regression of the relative difference N K R between buying and selling investors versus the relative sentiment indicator S R and the stock return Ret. The number in parentheses are the 5%-95% confidence intervals under Gaussian hypothesis and by using bootstrap analysis. The last two columns show the results of the partial correlation analysis. Investor α SR α Ret % variance of ρ(n K R, S R Ret) ρ(nk R, Ret S R ) category residual of N R K Companies.55 (.14,.95) -.61 (-.65,-.569) 63.3 % bootstrap (.15,.1) (-.685,-.548) Financial.18 (-.25,.62) -.52 (-.564,-.477) 73.1 % bootstrap (-.3,.64) (-.587,-.463) Governmental.21 (-.29,.71) (-.23,-.129) 96.8 % bootstrap (-.27,.75) (-.225,-.136) Non profit.25 (-.25,.75) (-.225,-.125) 96.9 % bootstrap (-.28,.79) (-.227,-.13) Households.68 (.26,.11) (-.68,-.523) 68.4 % bootstrap (.25,.111) (-.629,-.512) Foreign org..3 (-.17,.77) -.4 (-.446,-.353) 84.2 % bootstrap (-.15,.76) (-.449,-.354) 13 / 15

14 Comments The activity of governmental and non profit organizations is very poorly explained by return and news sentiment. Of the two factors, return plays clearly a major role. Households and companies are those for which sentiment and returns have the best explanatory power of their trading action. Return is clearly more important, but sentiment has also some explanatory power, especially when one consider the relative imbalance between buyers and sellers. For financial and foreign organizations the variance explained by the regressions is somewhat intermediate between the two pairs of categories above, but in general returns have a much higher explanatory power and sentiment plays a negligible role. For companies, financial institutions, households and foreign organizations α Ret < and large indicating that market polarization of trading actions is strongly anticorrelated with the Nokia return. The majority of single investors of these categories are therefore buying when the Nokia price goes down and selling when the price goes up. On a daily time scale, news move investors to trade. Most of the times the sentiment indicator is not significantly correlated with the imbalance between buyers and sellers. 14 / 15