Table 10: Results of regression analysis on AI affecting net income
(1) (2) (3) (4) (5) (6)
NI NI NI NI NI NI
AI 1783.1*** 223.0*** 106.0* 1778.0*** 26.27 15.58
(24.89) (5.01) (2.57) (24.78) (0.52) (0.33)
Size 80.82*** 29.50** -89.84
(7.96) (3.10) (-1.30)
Leverage 0.0404*** -0.00396** -0.0312
(113.63) (-2.66) (-0.67)
Age 5.214*** 2.316*
(4.68) (2.24)
Book-to-Market ratio 0.00681 0.00411 0.0145*
(0.10) (0.07) (2.32)
Operating Expenses 0.267*** -0.0682
(30.55) (-0.25)
Industry
522110.naics -556.2* 403.0
Commercial Banking (-2.19) (1.70)
522120.naics -527.9* 417.3
Savings Institutions (-2.06) (1.75)
522292.naics -472.0 428.5
Real Estate Credit (-1.12) (1.10)
522310.naics -513.9 417.6
Mortgage and Other Loan Brokers (-0.87) (0.77)
522390.naics -1073.6** -129.2
Other Activities Related to Credit
Intermediation (-2.93) (-0.38)
Ownership
0.stko 1072.7*** 705.1**
Publicly traded company (4.24) (3.02)
1.stko -238.7 -589.2*
Subsidiary of a publicly traded company
(-0.82) (-2.20)
3.stko 1182.8*** 753.2**
Company that is publicly traded but
not on a major exchange (4.64) (3.19)
Constant 34.55 -1213.4** -1380.1*** 34.97 178.7*** 1011.4
(1.69) (-3.22) (-3.96) (1.71) (51.80) (1.94)
Fixed Effect fyear gvkey +
fyear
gvkey + fyear
Clustering gvkey +
fyear gvkey + fyear
Observations 5553 5332 5332 5553 5518 5293
R2 0.100 0.773 0.807 0.101 0.911 0.925
t statistics in parentheses
33
* p < 0.05, ** p < 0.01, *** p < 0.001
Notes: The six regressions aim to regress the net income (NI) of the banks on the use of AI technologies with different control variables. Column (1) is a regression of net income on AI. Column (2) is a regression of net income on AI with the control variables size, industry, ownership, leverage, age, book-to-market ratio. Column (3) is a regression of net income on AI with the control variables size, industry, ownership, leverage, age, operating expenses, book-to-market ratio. Column (4) is a regression of net income on AI with year fixed effect. Column (5) is a regression of net income on AI with time and entity fixed effect and clustering. Column (6) is a regression of net income on AI with the control variables size, leverage, operating expenses, book-to-market ratio and with time and entity fixed effect and clustering.
The industry category “portfolio management” is used as base level. The ownership category “subsidiaries of companies that are not publicly traded” is used as base level.
Table 10 uses the net income of the bank as the dependent variable and AI as the dependent variable. The control variablesindustry, ownership, leverage, age, book-to-market ratio are added to regression (2), (3) and (6). The control variable operating expenses is added to regression (3) and (6). Specifications (1), (2), (3) do not use fixed effect nor clustering. The results are statistically significant at the 1% significance level for specification (1) and (2). Specification (3) is Specification (2) with the operating expenses added as a control variable. Since net income and operating expenses have a correlation coefficient of 0.895, adding the control to the model mitigates the effect. Specification (3) is significant at the 10% significance level. Specification (3) shows that net income is positively correlated with size, age and operating expenses but negatively correlated with leverage. The industry has a negligeable impact on net income as none of the coefficients are significant. The publicly traded companies (whether on a major exchange or not) have a higher net income than the subsidiaries of companies that are not publicly traded. The r-squared of specification (2) and (3) is high as the control variables explain properly the effect.
Specification (4) is specification (1) with time fixed effect and the AI coefficient remains positive and statistically significant at the 1% significance level. Specification (5) and (6) are not statistically significant due to the entity fixed effect and the few observations obtained regarding banks using AI. The coefficient of AI remains positive, specification (3) shows that banks using AI have a yearly net income higher by 106 million US dollars relative to banks not using AI. This evidence allows to expect that this positive effect of AI on Net Income would still be witnessed in a wider dataset.
Since Xiaodong Yuan, Fan Hou & Xuehui Cai (2020) used EBIT as a dependent variable, net income is replaced by EBIT in table 11. The results are similar except that specification (3) is statistically significant at the 1% significance level for the regression using EBIT as a dependent variable. Also, the coefficient for leverage is positive and significant at the 1% significance level. It is surprising since leverage should be more advantageous after taxes than before taxes however here we observe a positive coefficient when regressing EBIT on AI and a negative coefficient when regressing net income on AI.
34
Table 11: Results of regression analysis on AI affecting EBIT
(1) (2) (3) (4) (5) (6)
EBIT EBIT EBIT EBIT EBIT EBIT
AI 3989.2*** 369.8*** 153.7*** 3989.3*** 63.16 57.63
(25.97) (6.94) (3.58) (25.92) (1.44) (1.19)
Size 136.4*** 41.62*** 36.53
(11.22) (4.20) (1.19)
Leverage 0.0973*** 0.0154*** -0.0246
(228.89) (9.96) (-0.85)
Age 8.432*** 3.079**
(6.32) (2.86)
Book-to-Market ratio 0.0111 0.00614 0.00642
(0.14) (0.10) (0.80)
Operating Expenses 0.493*** 0.119
(54.19) (0.74)
Industry
522110.naics -2002.5*** -231.0
Commercial Banking (-6.58) (-0.94)
522120.naics -1964.0*** -218.4
Savings Institutions (-6.41) (-0.88)
522292.naics -1859.6*** -196.3
Real Estate Credit (-3.69) (-0.48)
522310.naics -1937.5** -217.1
Mortgage and Other
Loan Brokers (-2.75) (-0.38)
522390.naics -2386.9*** -642.6
Other Activities Related
to Credit Intermediation (-5.44) (-1.82)
Ownership
0.stko 1636.0*** 957.1***
Publicly traded company (5.40) (3.93)
1.stko 496.7 -150.5
Subsidiary of a publicly
traded company (1.43) (-0.54)
3.stko 1813.7*** 1020.3***
Company that is publicly traded but not on a major exchange
(5.94) (4.16)
Constant 82.82 -791.9 -1099.8** 82.81 405.2*** 142.0
(1.89) (-1.75) (-3.03) (1.89) (134.67) (0.67)
Fixed Effect fyear gvkey+fyear gvkey+fyear
Clustering gvkey+fyear gvkey+fyear
Observations 5547 5332 5332 5547 5512 5293
R2 0.108 0.930 0.955 0.109 0.990 0.990
t statistics in parentheses
* p < 0.05, ** p < 0.01, *** p < 0.001
Notes: The six regressions aim to regress the Earnings Before Interest and Taxes (EBIT) of the banks on the use of AI technologies with different control variables. Column (1) is a regression of EBIT on AI. Column (2) is a regression
35
of EBIT on AI with the control variables size, industry, ownership, leverage, age, book-to-market ratio. Column (3) is a regression of EBIT on AI with the control variables size, industry, ownership, leverage, age, operating expenses, book-to-market ratio. Column (4) is a regression of EBIT on AI with year fixed effect. Column (5) is a regression of EBIT on AI with time and entity fixed effect and clustering. Column (6) is a regression of EBIT on AI with the control variables size, leverage, operating expenses, book-to-market ratio and with time and entity fixed effect and clustering.
The industry category “portfolio management” is used as base level. The ownership category “subsidiaries of companies that are not publicly traded” is used as base level.
Figure 2: Net income of the banks using AI and the banks not using AI
Notes: The blue line represents the average net income per year of the 66 banks using AI from 2011 to 2019. The red line accounts for the average net income per year of the banks that do not use AI. Net income is expressed in million US dollars. Due to the scale, the average net income of the banks not using AI seems to be constant and equal to zero over the period. However, figure 4 shows that it is not the case.
36 Figure 3: Net income of the banks using AI
Notes: The blue line represents the average net income per year of the 66 banks using AI from 2011 to 2019. Net income is expressed in million US dollars.
Figure 4: Net income of the banks not using AI
Notes: The blue line represents the average net income per year of banks that do not use AI from 2011 to 2019. Net income is expressed in million US dollars.
37
Figure 5: Ratio of net income to total assets of the banks using AI and the banks not using AI
Notes: The blue line represents the average ratio of net income to total assets per year of the 66 banks using AI from 2011 to 2019.The red line represents the average ratio of net income to total assets per year of banks that do not use AI from 2011 to 2019. The average ratio of net income to total assets is expressed in percentages.
Figure 2 pictures the net income of the banks using AI and the banks not using AI over time, while figure 3 and 4 represents respectively the net income of the banks using AI and the net income the banks not using AI separately. Figure 5 shows the ratio of net income to total assets of the banks using AI and the banks not using AI over time. The net income of banks using AI is dramatically higher than the net income of banks not using AI (figure 2). It could partly be explained by reversed causality issue and omitted variable bias. However, when separating the lines in two different figures, it allows to display them with a more appropriate scale which exhibits that both types of banks have the same trend (figure 3 and figure 4). The net income of both types of banks increases dramatically from 2011 to 2019 probably because the banking sector is recovering from the subprime crisis of 2008 during this period. Nevertheless, banks using AI have a higher base level of around 1.6 billion dollars in 2011 against 15 million dollars for banks not using AI. The net income of banks using AI also grows at a higher rate than the net income of banks not using AI.
When scaling net income by total assets to control for the size of the bank (figure 5), the performance of the banks using AI is at least 1% higher than the performance of banks not using AI for the period 2011-2019. The maximum difference in performance reaches above 1.2% in 2012. Net income is not scaled by total assets in the previous regressions because the size is used as a control variable.
38
All AI coefficients are positive when using net income and EBIT as dependent variables. Thus, it seems that banks using AI have higher performance. The evidence supports the hypothesis stating that AI increases performance and it draws a similar conclusion as the paper from Amer Awad Alzaidi (2018). The following parts intend to understand the return drivers of banks using AI.