3. Data and Methodology
3.3 Tests of the Hypotheses
(bullish+1 and bearish-1), Y= return of TAIEX, Ai (i=1, 2,…,7) and Bi (i=1, 2,…,7) = regression coefficients , b and ai (i=1, 2,…,7) = intercept, ε= error term.
The regressions of the second topic are as below:
Model 3: S(t) = c1+ C1*lnR(t-1)+M(t)+M(t-1)+M(t-2)+M(t-3)+ ε….(8) S(t) = c2+ C2*lnR(t-2)+M(t)+M(t-1)+M(t-2)+M(t-3)+ ε….(9) S(t) = c3+ C3*lnR(t-3)+M(t)+M(t-1)+M(t-2)+M(t-3)+ ε….(10)
Where R= the number of responses, M= market return, S= return of a stock, Ci(i=1, 2, 3)= regression coefficients, ci (i=1, 2, 3) = intercept, ε= error term.
According to the regression above, the number of lags is three. However, not all the stocks have enough samples and the least required number of samples is ten so that the numbers of lags are not all three. Besides, we used the portfolio comprising the 23 stocks to run the regression. The regression is as below:
Model 4: PS(t) = c1+ C1*lnAR(t-1)+M(t)+M(t-1)+M(t-2)+M(t-3)+ ε….(11) PS(t) = c1+ C1*lnAR(t-2)+M(t)+M(t-1)+M(t-2)+M(t-3)+ ε….(12) PS(t) = c1+ C1*lnAR(t-3)+M(t)+M(t-1)+M(t-2)+M(t-3)+ ε….(13) Where AR= the average number of responses, M= market return, PS= return of the portfolio, Ci (i=1, 2, 3)= regression coefficients, ci (i=1, 2, 3) = intercept, ε= error term.
3.3 Tests of the Hypotheses
We will test our hypotheses using ordinary least squares (OLS) and use a significance level of 1%, 5% and 10% (two-tailed)
.
Before running the data, we delete the outliers of the data (1% of right tail and 1% of left tail) to avoid these data influencing the results.‧
4.1 Use the Atmosphere of PTT Stock Section to Predict the Return of TAIEX
In 3.2 (Research Methodology), we use the independent variable X representing the total number of bullish articles of specific stocks minus the bearish number and Y is the return of TAIEX. We assume that X represents the atmosphere of the stock market now and we want to know if this factor can predict the future return of TAIEX. The model 1 and 2 are used to exam the results. Results of model 1 and model 2 in chapter 3 are as below:
(p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
13
The number of samples for model 1 is 165 and the result of model 1 show that the atmosphere of PTT Stock Section can’t predict the return of TAIEX. Multicollinearity might exist among X(t-1), X(t-2),……, X(t-7) and then we use the model 2 in chapter 3. The result is as below:
Table 2: The result of model 2
Variable (1) lag1 (2) lag2 (3) lag3 (4) lag4 (5) lag5 (6)lag6 (7)lag7 Constant 0.00004742 0.0019 0.00342 0.00203 -0.00183 -0.00147 -0.00090736
(0.00239) (0.00250) (0.00255) (0.00265) 0.00258 0.00246 0.00238 lnX -0.0003884 -0.00099 -0.00148 -0.00103 0.000223 0.000106 -0.00007719
(0.00076) (0.00080) (0.00081) (0.00084) 0.00082 0.00079 0.00076
Observation 165 165 165 165 165 165 165
R square 0.00160 0.00930 0.01990 0.00900 0.00040 0.00010 0.00010 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The number of samples for model 2 is also 165 and the model 2 can avoid the multicollinearity. The result of model 2 shows there is significantly negative relation between lnX (t-3) and Y (t) (Table 2). It means the bullish atmosphere in stock section of PTT now will cause the negative return of TAIEX three weeks later and bearish atmosphere in stock section of PTT now will cause the positive return of TAIEX three weeks later. It means that the atmosphere in PTT delayed for three weeks.
‧
4.2 Use the number of responses of a specific stock to predict the return of stocks
We pick up 23 stocks from the component stocks of Taiwan 50 whose samples are at least 10. In ch3, we use the independent variable ( R ) to represent the number of responses in an article mentioning one of the 23 stocks – the author of article told us the stock is bullish or bearish. The dependent variable (S) is the returns of one of the 23 stocks next day or more days later. The model 3 in ch3 is used to exam whether the R can predict the S or not and the results are as below:
Table 3: The result of model 3 (ticker:4938)
Variable (8) lag1 (9) lag2 (10) lag3
The table 3 above shows that there is positive relationship between the number of responses and the returns 3 days from now of 4398.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
15
Table 4: The result of model 3 (ticker:3481)
Variable (8) lag1 (9) lag2 (10) lag3
Constant 0.013089 0.039979 -0.03759
(0.02480) (0.02416) (0.02383)
lnR -0.00337 -0.01336 0.015953*
(0.00927) (0.00881) (0.00884)
M(t) -0.0094 -0.01132* -0.0121*
(0.00708) (0.00641) (0.00631)
M(t-1) 0.004623 0.007167 0.005
(0.00593) (0.00575) (0.00544)
M(t-2) 0.005226 0.004176 0.002356
(0.00601) (0.00573) (0.00578)
M(t-3) 0.001646 0.002387 0.003797
(0.00593) (0.00563) (0.00560)
Observation 25 25 25
R square 0.1613 0.2466 0.2791 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 4 above shows that there is positive relationship between the number of responses and the returns 3 days from now of 3481.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
16
Table 5: The result of model 3 (ticker:2330)
Variable (8) lag1 (9) lag2 (10) lag3
Constant -0.0011 0.011279 0.004166
(0.00704) (0.00697) (0.00747)
lnR 0.000717 -0.00444 -0.00149
(0.00276) (0.00274) (0.00298)
M(t) 0.001833 0.00061 0.001459
(0.00297) (0.00293) (0.00309)
M(t-1) 0.00348 0.00434 0.00300
(0.00313) (0.00301) (0.00312)
M(t-2) -0.00384 -0.00507 -0.00362
(0.00311) (0.00305) (0.00314)
M(t-3) 0.004675 0.004297 0.004293
(0.00310) (0.00297) (0.00320)
Observation 35 35 35
R square 0.1498 0.2183 0.1551 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 5 above shows that there is no significant relationship between the number of responses and returns of 2330.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
17
Table 6: The result of model 3 (ticker:3008)
Variable (8) lag1 (9) lag2 (10) lag3
Constant 0.015061 0.007836 0.015109
(0.00920) (0.00941) (0.00879)
lnR -0.00556 -0.00242 -0.0057
(0.00368) (0.00377) (0.00357 )
M(t) 0.000239 -0.0007 -0.00105
(0.00361) (0.0037) (0.00358)
M(t-1) 0.00043 0.00003 -0.00062
(0.00369) (0.00381) (0.00365)
M(t-2) 0.00020 0.00140 0.00191
(0.00376) (0.00386) (0.00374)
M(t-3) -0.00053 -0.00163 -0.00063
(0.00365) (0.00376) (0.00363)
Observation 36 36 36
R square 0.0763 0.0196 0.0841 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 6 above shows that there is no significant relationship between the number of responses and returns of 3008.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
18
Table 7: The result of model 3 (ticker:2303)
Variable (8) lag1 (9) lag2
Constant -0.01268 0.00817
(0.01373) (0.00899)
lnR -0.0028 -0.0158**
(0.01014) (0.00538)
M(t) 0.02786 0.02656*
(0.02123) (0.01328)
M(t-1) -0.00755 -0.00852
(0.02096) (0.00896)
Observation 10 10
R square 0.2694 0.6963 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 7 above shows that there is negative relationship between the number of responses and the returns 2 days from now of 2303.
‧
Table 8: The result of model 3 (ticker:2357)
Variable (8) lag1 (9) lag2 (10) lag3
Constant -0.01248 -0.00272 0.004495
(0.01166) (0.01229) (0.01118)
The table 8 above shows that there is no significant relationship between the number of responses and returns of 2357.
Table 9: The result of model 3 (ticker:1303) Variable (8) lag1
The table 9 above shows that there is no significant relationship between the number of responses and returns of 1303.
‧
Table 10: The result of model 3 (ticker:2002)
Variable (8) lag1
(p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 10 above shows that there is no significant relationship between the number of responses and returns of 2002.
Table 11: The result of model 3 (ticker:2311)
Variable (8) lag1 (9) lag2 (10) lag3
The table 11 above shows that there is negative relationship between the number of responses and the returns 3 days from now of 2311.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
21
Table 12: The result of model 3 (ticker:2317)
Variable (8) lag1 (9) lag2 (10) lag3
Constant 0.008615* -0.01164** -0.00199
(0.00510) (0.00483) (0.00526)
lnR -0.003 0.005511*** 0.00145
(0.00202) (0.00191) (0.00206)
M(t) 0.00634** 0.00840*** 0.00669**
(0.00250) (0.00241) (0.00254)
M(t-1) 0.00503* 0.00670*** 0.00621**
(0.00262) (0.00243) (0.00269)
M(t-2) 0.00283 0.00400 0.00296
(0.00263) (0.00250) (0.00271)
M(t-3) 0.004479* 0.00469* 0.005571**
(0.00261) (0.00241) (0.00267)
Observation 48 48 48
R square 0.2567 0.3476 0.2270 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 12 above shows that there is positive relationship between the number of responses and the returns 2 days from now of 2317.
‧
Table 13: The result of model 3 (ticker:1216)
Variable (8) lag1 (9) lag2 (10) lag3
The table 13 above shows that there is significantly positive relationship between the number of responses and the returns 1 day from now of 1216.
Table 14: The result of model 3 (ticker:2382) Variable (8) lag1
The table 14 above shows that there is no significant relationship between the number of responses and returns in 2382.
‧
Table 15: The result of model 3 (ticker:2891)
Variable (8) lag1 (9) lag2
The table 13 above shows that there is significantly negative relationship between the number of responses and the returns 1 day from now of 2891.
Table 16: The result of model 3 (ticker:2325)
Variable (8) lag1 (9) lag2 (10) lag3
The table 16 above shows that there is significantly positive relationship between the number of responses and the returns 1 day from now of 2325.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
24
Table 17: The result of model 3 (ticker:2412)
Variable (8) lag1 (9) lag2
Constant -0.00841 -0.00205
(0.00718) (0.00531)
lnR 0.00291 0.00071
(0.00252) (0.00204)
M(t) 0.00558*** 0.00549***
(0.00071) (0.00079)
M(t-1) 0.00054 0.00090
(0.00084) (0.00091)
M(t-2) 0.00029 0.00039
(0.00074) (0.00082)
Observation 10 10
R square 0.9325 0.9164 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 17 above shows that there is no significant relationship between the number of responses and returns in 2412.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
25
Table 18: The result of model 3 (ticker:2490)
Variable (8) lag1 (9) lag2 (10) lag3
Constant 0.00981 -0.01012 0.02060
(0.0241) (0.0266) (0.0209)
lnR -0.00535 0.00296 -0.01056
(0.0096) (0.0107) (0.0088)
M(t) 0.00405 0.00381 0.00450
(0.0051) (0.0058) (0.0048)
M(t-1) -0.00360 -0.00257 -0.00057
(0.0050) (0.0052) (0.0051)
M(t-2) -0.00133 -0.00182 -0.00428
(0.0053) (0.0053) (0.0051)
M(t-3) -0.00491 -0.00342 -0.00318
(0.0088) (0.0084) (0.00797)
Observation 20 20 20
R square 0.1161 0.1015 0.1800 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 18 above shows that there is no significant relationship between the number of responses and returns in 2490.
‧
Table 19: The result of model 3 (ticker:1476)
Variable (8) lag1 (9) lag2 (10) lag3
Constant -0.00219 -0.02264* 0.008389
(0.01295) (0.01305) (0.01101)
The table 19 above shows that there is significantly negative relationship between the number of responses and the returns 3 days from now of 1476.
Table 20: The result of model 3 (ticker:9904) Variable (8) lag1
The table 20 above shows that there is no significant relationship between the number of responses and returns of 9904.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
27
Table 21: The result of model 3 (ticker:2354)
Variable (8) lag1 (9) lag2 (10) lag3
Constant 0.011796 0.03279 0.010732
(0.02871) (0.02670) (0.03119)
lnR -0.00871 -0.01894 -0.00761
(0.01462) (0.01265) (0.01513)
M(t) 0.00018 -0.001 0.003516
(0.01012) (0.00891) (0.01137)
M(t-1) 0.005509 -0.00047 0.00188
(0.01076) (0.00966) (0.01135)
M(t-2) 0.009579 0.015187 0.010182
(0.01103) (0.00944) (0.01095)
M(t-3) 0.005619 -0.00245 0.002736
(0.01203) (0.00947) (0.01049)
Observation 12 12 12
R square 0.3181 0.4740 0.3069 (p-value<0.01: ***, p-value < 0.05: **, p-value < 0.1: *)
The table 21 above shows that there is no significant relationship between the number of responses and returns of 2354.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
28
Table 22: The result of model 3 (ticker:2474)
Variable (8) lag1 (9) lag2 (10) lag3
Constant -0.01403 0.015708 0.010182
(0.01408) (0.01497) (0.01534)
lnR 0.005135 -0.00986 -0.00704
(0.00645) (0.00711) (0.00729)
M(t) 0.01809*** 0.015336*** 0.016977***
(0.00524) (0.00532) (0.00519)
M(t-1) 0.002762 0.000415 0.000228
(0.00537) (0.00524) (0.00548)
M(t-2) -0.00061 -0.0004 -0.00118
(0.00593) (0.00582) (0.00592)
M(t-3) -0.0144** -0.01282** -0.01339**
(0.00626) (0.00612) (0.00619)
Observation 37 37 37
R square 0.3846 0.4087 0.3904 (p-value<0.01: ***, p-value < 0.05: **, p-value < 0.1: *)
The table 22 above shows that there is no significant relationship between the number of responses and returns of 2474.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
29
Table 23: The result of model 3 (ticker:2454)
Variable (8) lag1 (9) lag2 (10) lag3
Constant 0.01273 -0.01463 0.01702
(0.0124) (0.0115) (0.0128)
lnR -0.00244 0.01013** -0.00457
(0.0052) (0.00491) (0.0056)
M(t) -0.00225 -0.00485 -0.00110
(0.0062) (0.0059) (0.0062)
M(t-1) 0.00115 0.00180 0.00214
(0.0059) (0.00546) (0.0059)
M(t-2) -0.00406 -0.00444 -0.00417
(0.0058) (0.00535) (0.0057)
M(t-3) -0.00485 -0.00517 -0.00410
(0.0061) (0.0056) (0.00598)
Observation 31 31 31
R square 0.0557 0.1862 0.1862 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 23 above shows that there is significantly positive relationship between the number of responses and the returns 2 days from now of 2454.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
30
Table 24: The result of model 3 (ticker:2887)
Variable (8) lag1 (9) lag2 (10) lag3
Constant -0.01603 0.011135 0.000237
(0.01135) (0.00803) (0.00859)
lnR 0.00598 -0.00395 -0.00002
(0.003875) (0.002458) (0.00283)
M(t) 0.01483*** 0.015001*** 0.01466***
(0.00162) (0.00161) (0.00196)
M(t-1) -0.00239 -0.00072 -0.00081
(0.00207) (0.001778) (0.0022)
M(t-2) -0.00170 -0.00036 -0.00192
(0.00197) (0.00217) (0.00239)
M(t-3) -0.00115 0.00673 0.00252
(0.00733) (0.00733) (0.00852)
Observation 11 11 11
R square 0.9577 0.9587 0.9375 (p-value<0.01: ***, p-value < 0.05:**, p-value < 0.1: *)
The table 24 above shows that there is no significant relationship between the number of responses and returns of 2887.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
31
Table 25: The result of model 3 (ticker:3474)
Variable (8) lag1 (9) lag2 (10) lag3
Constant 0.013083 0.002841 0.027658*
(0.01557) (0.01552) (0.01480)
lnR -0.00557 -0.00144 -0.0114*
(0.006019) (0.005981) (0.00568)
M(t) 0.01059** 0.01034** 0.01062***
(0.00401) (0.004121) (0.00389)
M(t-1) -0.00336 -0.00206 -0.00348
(0.00429) (0.00412) (0.00401)
M(t-2) 0.007* 0.00692 0.007935*
(0.00408) (0.004295) (0.00396)
M(t-3) -0.00254 -0.00163 -0.00403
(0.00414) (0.004058) (0.004068)
Observation 55 55 55
R square 0.1674 0.1538 0.2172 (p-value<0.01: ***, p-value < 0.05: **, p-value < 0.1: *)
The table 25 above shows that there is significantly negative relationship between the number of responses and the returns 3 days from now of 3474.
‧
Table 26: The result of model 4 (portfolio)
Variable (11) lag1 (12) lag2 (13) lag3
The table 26 above shows the portfolio that there is significantly positive relationship between the average number of responses and the returns 1 day from now and negative relationship 3 days from now.
According to the Table 3 to Table 25, we can find the independent variables of stocks 4938, 3481, 2303, 2311, 2317, 1216, 2891, 2325, 1476, 2454 and 3474 can predict the return of these stocks and the numbers of lags significant for all stocks above are not all the same. The relationship between independent and dependent variables can be positive or negative. Even if some of the independent variables can predict the dependent variables, the results are inconsistent. Besides, the table 26 shows the portfolio that there is significantly positive relationship between the average number of responses and the returns 1 day from now and negative relationship 3 days from now.
We also can not explain why the significant numbers of lags are 1 and 3 but 2.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
33
Consequently, the information can not give us consistent results. We can’t explain rationally about these results and we can say that the people in PPT are just a specific group. Their comments or articles in PTT aren’t able to influence the stock market. The information is insufficient.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
34
5. Conclusion
This paper use the social network PTT to study if the articles of stock section of PTT can predict the return of TAIEX and specific stocks. According to the model 2, the result is if the atmosphere in PTT is bullish/bearish this week, the return of TAIEX will be negative/positive three weeks from now. However, it is hard to explain why the result doesn’t show that other independent variables are significant but the X (t-3).
Consequently, the information isn’t enough so that we can’t explain the result.
According to the model 3, the components of Taiwan 50 like 4938, 3481, 2303, 2311, 2317, 1216, 2891, 2325, 1476, 2454 and 3474 can use the numbers of responses of article to predict their returns few weeks from now. However, not only some of the relations between independent and dependent are positive and other are negative but also the numbers of lags significant for every stock are different. These information can’t give us consistent results. The model 4 shows the portfolio that there is significantly positive relationship between the average number of responses and the returns 1 day from now and negative relationship 3 days from now. We also can not explain why the significant numbers of lags are 1 and 3 but 2. We can’t explain rationally about these results and we can say that the people in PPT are just a specific group. Their comments or articles in PTT are not able to influence the stock market.
The information is insufficient.
‧
1. Eugene Fama, 1970, Efficient capital markets: A review of theory and empirical work, The Journal of Finance 25, 383-417.
2. Kahneman and Tversky, 1979, Prospect Theory-Division of the Humanities and Social Science, Econometrica 47, 263-292.
3. Wesley S. Chan, 2002, Stock Price Reaction to News and No-News: Drift and Reversal After Headlines, Journal of Financial Economics 70, 223–236.
4. Werner Antweiler and Murray Z. Frank, 2004, Is All That Talk Just Noise? The Information Content of Internet Stock Message Boards, The Journal of Finance 3, 1259-1294.
5. Malcolm Baker and Jeffrey Wurgler, 2006, Investor Sentiment and the Cross-Section of Stock Returns, The Journal of Finance 4, 1645-1680.
6. Z. Da, J. Engelberand, and P. Gao, 2010, The sum of all fears: investor sentiment and asset prices, http://ssrn.com/abstract=1509162
7. Johan Bollen, Huina Mao and Xiao-Jun Zeng, 2010, Twitter mood predicts the stock market, Indiana University, Computer 1010, 1–8.
8. Huina Mao, Scott Counts and Johan Bollen, 2011, Predicting Financial Markets:
Comparing Survey, News, Twitter and Search Engine Data, arXiv.org, Quantitative Finance Papers, 1051-1112
9. Tushar Rao and Saket Srivastava, 2012, Analyzing Stock Market Movements Using Twitter Sentiment Analysis, IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining.
10. Eric D. Brown, 2012, Will Twitter Make You a Better Investor? A Look at sentiment, User Reputation and Their Effect on the Stock Market, Dakota State University