1

i have a large dataset (~3,000 datapoints, 6H interval) on Twitter and Bitcoin Data and try to estimate the effect of tweets on price changes / trading volume of Bitcoin.

Therefore, i run a VAR model with 3 lags as a result of ACFs and AIC/BIC and get the following output:

Estimation results for equation TradingV: 
========================================= 
TradingV = LM.l1 + TweetV.l1 + Price.l1 + TradingV.l1 + LM.l2 + TweetV.l2 + Price.l2 + TradingV.l2 + LM.l3 + TweetV.l3 + Price.l3 + TradingV.l3

         Estimate Std. Error t value             Pr(>|t|)    



LM.l1       -0.285972   0.076378  -3.744             0.000185 ***
TweetV.l1   -0.007748   0.072017  -0.108             0.914329

Price.l1    -2.267196   0.564289  -4.018            0.0000603 ***
TradingV.l1 -1.119082   0.019625 -57.022 < 0.0000000000000002 ***
LM.l2       -0.123553   0.078909  -1.566             0.117512

TweetV.l2    0.103206   0.080115   1.288             0.197771

Price.l2    -1.608752   0.639476  -2.516             0.011933 *

TradingV.l2 -0.942809   0.023328 -40.415 < 0.0000000000000002 ***
LM.l3        0.015644   0.075695   0.207             0.836281

TweetV.l3    0.024549   0.071821   0.342             0.732523

Price.l3    -0.958493   0.565993  -1.693             0.090474 .

TradingV.l3 -0.439742   0.019386 -22.683 < 0.0000000000000002 ***




Signif. codes:  0 ‘*’ 0.001 ‘’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1


Residual standard error: 0.5044 on 2876 degrees of freedom
Multiple R-Squared: 0.611,  Adjusted R-squared: 0.6093 
F-statistic: 376.4 on 12 and 2876 DF,  p-value: < 0.00000000000000022


Covariance matrix of residuals:
                 LM      TweetV       Price   TradingV
LM        0.0161318 -0.00123492  0.00036625 -0.0142670
TweetV   -0.0012349  0.01898572  0.00008595  0.0334891
Price     0.0003663  0.00008595  0.00025976 -0.0009738
TradingV -0.0142670  0.03348907 -0.00097377  0.2544255


Correlation matrix of residuals:
               LM   TweetV   Price TradingV
LM        1.00000 -0.07056  0.1789  -0.2227
TweetV   -0.07056  1.00000  0.0387   0.4818
Price     0.17891  0.03870  1.0000  -0.1198
TradingV -0.22269  0.48185 -0.1198   1.0000

As you can see, Tweet Volume (TweetV) does not seem to have a statistical significant influence on trading volume (TradingV). However, a Granger causality test suggests that TweetV does Granger-cause TradingV and the IRF(shock in TweetV and response of TradingV) also suggests a statistical significant influence:

IRF with a shock in TweetV, response of TradingV

How is that possible, or what could be a possible explanation for this?

Thanks in advance.

Dalogh
  • 11
  • 1
    Both the GC and the IRF are functions of all lags. In, e.g., the case of GC, we test if all lags have zero coefficients via an F-test, and (you can find examples on this site, too, e.g. https://stats.stackexchange.com/questions/151403/significance-of-individual-coefficients-vs-significance-of-both/151410#151410) it is possible that an F-test rejects while all t-tests do not. – Christoph Hanck Nov 28 '22 at 12:41

0 Answers0