The Effect of Digital Platform Strategies on Firm Value in the Banking Industry

ABSTRACT After digital platforms have become successful in the information technology (IT) industry, incumbents from traditional industries increasingly implement digital platform strategies. However, there is mixed evidence on whether these incumbents benefit from digital platform strategies. To provide systematic insights, we focus on the banking industry. With the emergence of open banking, banks have begun implementing digital platforms to unlock the innovative power of third-party developers. We conducted an event study based on the announcement of digital platform strategies in a global sample of 165 banks. We show that, on average, investors react positively to the announcements. Contrary to our expectations, this effect is more substantial for banks from emerging markets than those from developed markets. Prior artificial intelligence (AI) orientation only partly contributes to investors’ favorable perception of a digital platform strategy. These results point to the complex interplay of AI orientation and digital platform strategies, yielding questions for future research.


List of Countries of Origin
: developed markets are part of the "High-Income Economies," emerging markets are part of the "Higher-Middle-Income Economies" or "Lower-Middle-Income Economies")

Details on the Event Study Analysis
As we ensured that all banks were available in the Bloomberg database, we used the Bloomberg Excel add-in to obtain the respective stock prices. 1After calculating returns from the banks' stock prices, we used the market model based on the capital asset pricing model (CAPM) to calculate the banks' abnormal returns [6,8] 2 .For this purpose, we estimated the following Ordinary Least Squares (OLS) model: where   is the return for company  on day ,   is the return of the market index  on day  and   is the error term.As a market index, we used the MSCI World Index, also obtained from Bloomberg (ticker: MXWO Index), for the main analysis. 3o avoid a bias in the parameter estimations around the event date, we estimated the parameters using the market model with a 200-trading-day window ending 20 days before the event day [7].
In the next step, we used the estimated parameters  �  and  ̂ from the market model in equation (1) and calculated abnormal returns   for each bank as follows: We then averaged the individual banks' abnormal returns   to receive the average abnormal return   for each day : where  is the number of events (and, in our case, the number of banks).To finally calculate cumulative average abnormal returns ([ 1 ,  2 ]) for a specific event window that ranges from  1 to  2 , we summed up the respective   that are included in the event window: We measured cumulative average abnormal returns over the event windows , and [-3,3].For all analyses, we used Stata's estudy package [27].The package includes the calculation of cumulative average abnormal returns and the corresponding test statistics.We report the t-test [22], the Patell test [28,29], the BMP test [4], and the KP-adjusted version of both tests, following Kolari and Pynnonen [21].We report additional parametric tests beyond the t-test to address limitations of the t-test related to event-induced volatility and crosssectional correlation of abnormal returns.
The basic idea of parametric tests is to derive a test statistic that follows a specific distribution (e.g., standard normal distribution in the case of the t-test when N is large).Implementing such a test typically requires the standard deviation of the cumulative average abnormal return 2 is unknown, it is appropriate to use an estimate based on the estimation window [ 1 ,  2 ] where, as outlined above,  1 = −220 and  2 = −20 [22].Hence, we calculated the standard deviation of the abnormal return using the daily abnormal returns from the estimation window (

2
. For event windows longer than one day, we 2 by multiplying it by length of the event window.For the t-test, for instance, the test statistic is then derived by where the calculation of  [ 1 , 2 ] is depicted in equation ( 5) above.This test statistic will asymptotically follow a standard normal distribution when N is large so we can easily derive its statistical significance.For details on the test statistics of the Patell test [28,29], the BMP test [4], and their adjusted versions [21], we refer to the respective original sources.
As an alternative to CAPM for calculating abnormal returns, the Fama-French three-factor model (FFM) has been suggested [15] because it captures market risks better.We were unable to use the Fama-French three-factor model for our event study because the required correction factors (size correction factor (SML) and book-to-market correction factor (HML)) were only available for developed markets on a daily basis [16].However, we used the Fama-French model for the subsample of companies from developed markets as a robustness test for the results obtained based on the CAPM (see Table 4 in the main article).For this purpose, we extend equation (1) above by two factors and estimate the following OLS model: where   is the return for company  on day ,   is the risk-free rate of the turn (US treasury bills) on day ,   is the return of the market index  on day ,   is the size correction factor for day ,   is the book-to-market correction factor for day , and   is the error term.
The factors for the Fama-French model were obtained from Fama and French [16].Parameter estimation was again done with a 200-trading-day window ending 20 days before the event day to avoid any bias caused by the events themselves [7].
Then, the abnormal returns   were calculated using equation ( 8), as shown below.Taking the parameters  �  ,  ̂,  �  , and  ̂ from the results of the regression based on equation (7).
Next, we calculated cumulative average abnormal returns (CAAR) in the same way as in our main analysis, as shown in equations ( 3) and ( 4).Based on the CAAR values, we performed the same battery of parametric tests using Stata's estudy package [27].The results are included in Table 4 in the main article, which shows the analysis of the moderating effects.

Robustness Test -Self-Selection Bias
As a digital platform strategy announcement may not be random but determined by specific bank characteristics, our findings might suffer from a potential self-selection bias.To control for this potential bias, we apply-in line with existing studies [6,11,23]-a Heckman correction [19].
In the first stage of the Heckman correction, we model the probability of a bank announcing a digital platform strategy.For this purpose, we include the share of other banks in the sample that have already announced a digital platform strategy because it is likely that similar activities from peers affect the focal bank's decision; we refer to this as peer effect.While Wolfolds and Siegel [35] emphasize the careful use of instruments in the Heckman correction, similar identification strategies have been applied by Fotheringham and Wiles [17] in the context of chatbot launches and by Bhagwat, Warren, Beck and Watson [3] in the context of corporate activism.Importantly, such a peer variable satisfies the exclusion restriction as it only has an effect on the probability of the firm's action (i.e., announcing a digital platform strategy) but not on the investors' evaluation of a specific announcement (i.e., abnormal returns after announcing a digital platform strategy).
We furthermore expect large banks to be more likely to implement a digital platform strategy.
Therefore, we included total assets as a proxy for size as an independent variable in the first stage of the Heckman correction.To capture all missing variables related to the specific year of implementation, we included year-fixed effects and consequently estimated the following probit model: where Probability(  = 1) is a binary variable, being one if bank  has announced a digital platform strategy in year  and zero in the years before.In line with similar studies [e.g., 17], we use the log-transformed total assets of bank  from the previous year (i.e.,  − 1).As we expect differences between banks operating in developed markets and banks operating in emerging markets, we calculate peer effects depending on the market development status .Thus,   is defined as the percentage of banks in the developed (emerging) sample that have announced a digital platform strategy in year .Finally,   represents the error term.
After estimating equation ( 9), we calculated the inverse Mill's ratio λ  for every bank  in year  as follows: where Χ represents a vector that includes the independent variables of equation ( 9) above.(•) denotes the standard normal probability density function and Φ(•) the corresponding cumulative distribution function.
For the second stage of the Heckman correction, the inverse Mill's ratio serves as an independent variable to explain the cumulative abnormal returns of the event.Hence, we estimate a simple cross-sectional regression as there is only one observation in the event year for each bank .
Following previous studies [e.g., 5], this cross-sectional analysis uses a standard OLS regression with cumulative standardized abnormal returns (  ) as the dependent variable.More specifically, we estimate the following OLS regression model: Although this step is only used to calculate the respective inverse Mill's ratios (i.e., λ  ) for the second stage, it is interesting to observe that the peer effects significantly drive a bank's probability of announcing a digital platform strategy.Columns 2-5 display the results for the second stage of the Heckman correction.The insignificant coefficients for λ  across all event windows suggest that our findings do not suffer from a potential self-selection bias. 4Note that the Heckman correction assumes a bivariate normal distribution of the error terms in both stages; thus, we excluded eight outliers to comply with this assumption.Moreover, the Bloomberg database did not provide total assets for one bank in our sample, and we consequently had to exclude this bank.Therefore, the number of banks for this exercise decreases from 165 to 156.Selection concerns in event studies have also been discussed in finance by Ahern [2].The author shows that the statistical tests for event studies from a non-random sample can be biased.This bias occurs when firms in the sample are grouped by specific characteristics, for example, size, accounting ratios, or prior returns.Ahern [2] recommends using a so-called "characteristic-based benchmark model" instead of the classical market index to remedy this bias.The idea of this approach is that the underlying index resembles the firms analyzed in the event study.Since our sample might also represent a grouped sample as banks are similarly exposed to economic shocks (e.g., interest rates, macroeconomic news, etc.), we re-estimated the event study for the main effect hypothesized in Hypothesis 1 using an index that only comprises banks and related financial institutions (i.e., MSCI World Financials Index, Bloomberg ticker MXWO0FN Index).
The results for this additional robustness test are shown in Table 2.5 and are qualitatively similar to the main analysis with the MSCI World Index, albeit less significant.

Robustness Test -Cross-Sectional Regression on Abnormal Returns
In the main analysis, we assessed the moderating effects by comparing the cumulative abnormal average returns of the subsamples.To jointly examine the effects of the subsamples, we apply a cross-sectional regression analysis in which we regress the hypothesized moderating variables (developed market and AI orientation) as well as bank-specific control variables on cumulative standardized abnormal returns (  ).The estimation equation is specified as follows: where   is the cumulative standardized abnormal return for bank  during the event as calculated in Bose and Leung [6].For this analysis, the variables of interest are   and   :   is a dummy variable being one if bank  is in the developed market subsample and zero otherwise.  is a dummy being one if bank  is in the AI orientation subsample and zero otherwise.We refer to Table 2.6 for the variable description for the bank-specific control variables.We include event-year dummies to account for unobserved year-specific effects in the abnormal returns and use robust standard errors.  represents the remaining error term.We report the results for a quantile regression as applied in Bose and Leung [6] and for a standard OLS regression.The results of the cross-sectional regression analysis are shown in Panel A of Table 2.7 for the event windows [0,2] and [-3,3], respectively.While the moderating effect of developed markets is insignificant in both event windows, we observe a significant moderating effect for AI orientation in the event window [0,2] for one specification.Hence, there is some evidence that AI orientation has a significant moderating effect.Note that the Bloomberg database did not provide total assets for one bank in our sample (see also for the Heckman correction above) and, in addition, other financial control variables for one more bank in our sample.Consequently, we reduced the sample size from 165 to 163. 5lthough the inclusion of bank-specific control variables in the regression analysis accounts for differences in bank characteristics, existing literature commonly applies propensity score matching to further reduce the risk of estimation biases that arise from such differences [e.g., 26,33].Consequently, we also applied a propensity score matching before the cross-sectional regression for additional robustness.We use kernel matching based on bank-specific variables [e.g., 18,25] and apply the propensity score matching for both subsamples separately.After matching, the average reduction of standardized bias of covariates for developed market and AI orientation are 50.3%and 71.7%, respectively.This reduction indicates that the matching procedures produced a better balance, yielding less bias in treatment effects estimation.
First, we created a matched sample for the development market subsample and estimated equation ( 12) above (Panel B of Table 2.7) 6 .The matching reduces our sample size to 141 (i.e., the excluded banks were too different to be matched with a counterpart).With the matched sample, a significant moderating effect of developed market emerges in the event window [-3,3] for one specification, and one more coefficient is at the brink of statistical significance (i.e.,  = 0.108 for [0,2] in OLS).A nonparametric Wilcoxon rank-sum test further supports the effect for both event windows.The test shows a statistically significant difference (at the 10% level) between the standardized cumulative abnormal returns of the two matched groups of banks from developed vs. emerging markets.
Second, we created a matched sample for the AI orientation subsample and estimated equation ( 12) above (Panel C of Table 2.7).This matching reduces our sample size only slightly to 157 banks.This matching confirms the significance of the moderating effect of AI orientation for the event window [0,2].Again, the effect is also identified as statistically significant at the 10% level with a Wilcoxon rank-sum test.

Table 1 . 3 .
List of Countries of Origin (classification based on The World Bank's Country and Lending Groups Classification

Table 2 . 1 .
Analysis of the Main Effect for Event Windows Before the Announcement

Table 2 .
2. Analysis of the Main Effect with Stock Prices in Local CurrenciesRobustness

Table 2 .
3. Analysis of the Main Effect with the MSCI ACWI Index as the Market Index Bose and Leung (2019)+  4  +  (11)where   is the cumulative standardized abnormal return for bank  as calculated in, for example,Bose and Leung (2019).λ  represents the inverse Mill's ratio for every bank  of the event year.Notably, a statistically significant coefficient for λ  would indicate a potential selfselection bias.  is a dummy variable being one if bank  is in the developed market subsample and zero otherwise.  is a dummy being one if bank  is in the AI orientation subsample and zero otherwise.We additionally include event-year dummies to account for unobserved year-specific effects in the abnormal returns.We use robust, nonclustered (as we only have one observation per bank) standard errors.againrepresentsthe remaining error term.The results of this analysis are shown in Table2.4.Column 1 presents the first-stage output.

Table 2 .
5. Analysis of the Main Effect with the MSCI World Financials Index Natural logarithm of age in years for bank .ln(  ) Natural logarithm of total assets (Bloomberg field BS_TOT_ASSET) for bank  in the event year.ln(1+)Natural logarithm of return on assets (Bloomberg field RETURN_ON_ASSET) for bank  in the event year.Since few banks have a negative return on assets and the logarithm cannot be computed for negative values, we add a constant to all values in the sample [e.g., 34].ln(  )Natural logarithm of the ratio debt (Bloomberg field SHORT_AND_LONG_TERM_DEBT) to equity (Bloomberg field TOTAL_EQUITY) for bank  in the event year.