Which Banks Recover from a Banking Crisis? Preliminary and incomplete conference version Usual disclaimer applies

1. Which Banks Recover from a Banking Crisis? Preliminary and incomplete conference versionUsual disclaimer applies Emilia Bonaccorsi di Patti Bank of Italy Anil K Kashyap University of Chicago, Graduate School of Business Federal Reserve Bank of Cleveland-FDIC Joint Financial Stability Conference, April 17, 2008

2. Motivation: Policy • Empirically: what distinguishes banks which do and do not recover from sudden declines in profitability? • Why care? • Collapses in bank health have become common place • Disagreements about the best way for regulators to proceed during a crisis: PCA (limit asset growth, raise capital) versus adjustments to portfolio, other?

3. This Paper • Analyze bank-level data to determine factors that govern recovery from distress. • Employ matched bank and borrower data to study how lending policies change after the shock to compare recovering and non-recovering banks Limitations: • Take the (un-modeled) initial shock to profits as given • Can’t say anything about how to avoid the shocks or what they are But…..given the paucity of existing evidence on what does and does not matter, a reduced form analysis is a useful starting point

4. Goals Three questions: • What are the key differences in the characteristics of the banks that do and do not recover? • How much of the recovery depends on macro or regional conditions that are out of the hands of the individual banks? • To the extent that bank choices do matter, which ones are most important and why?

5. Main Findings • Banks that get into trouble were lending to riskier clients than the average in the overall economy. Both have higher costs than their peers but costs are unrelated to recovery • Recovery depends both on factors that banks can and cannot control: • Size of the initial profit drop that occurs at the onset of distress • General business climate after the shock also matters, • Adjustments made by the bank in the wake of the shock: ability to adjust the loan portfolio and reduce default rate is critical • Matched loan/bank data: recovering banks do NOT make across the board reductions in credit -- if anything they are more likely to keep lending • BUT, if the share of high risk customers is substantial, the recovering banks show more of a propensity to tighten credit

6. Background Data

7. The Sample of Distressed Banks • All banks operating in Italy excluding: • i) cooperative banks, • ii) foreign bank branches, • iii) banks with assets<51 million euros as of 1995, • iv) banks with charter less than four years old • A bank is considered distressed in year t if: i) ROA (profits before tax/GTA) drops by at least 50% ii) the bank moves from above to below the 25th percentile of the distribution of ROA • 31 cases where a bank meets these conditions more than once  count only first episode • 151 banks distressed 1987-2004  to allow for sufficient leads and lags only banks with crisis between 1989-2001: sample of 120 banks out of 250/year

8. The Sample of Distressed Banks

9. Recovery • Recovery is a combination of improved performance and persistence of improvement • A bank is considered to have recovered if any holds: 1) At t = 1 its ROA is greater than 25th percentile and at t=2 ROA is greater or equal to ROA the year before the shock or the ROA percentile is greater or equal to the percentile observed the year before the shock; or 2) At t = 2 its ROA is greater than the 25th percentile and at t=3 ROA is greater or equal to ROA the year before the shock, or the ROA percentile is greater or equal to the percentile observed in the year before the shock; or 3) At t = 3 ROA is greater than the 25th percentile and at t=4 ROA is greater or equal to ROA the year before the shock, or the ROA percentile is greater or equal to the percentile observed in the year before the shock -41 recovering banks

10. Recovering Banks

11. Logit Regressions on Bank-level Data • Recover=1, 0 otherwise, conditional on survival for at least 1 year after shock 1. Macroeconomic factors: recovery is pre-determined and depends on external conditions: regional dummies for North and South (Center is excluded) and regional GDP growth in post crisis years 2. Bank-Specific Factors: • Recovery is determined by the size of the initial profit decline (use ROA relative to the average value of ROA for banks of same size in the same region) • Banks take actions to improve  R/NR banks tend to differ in the default rate after the shock: add the default rate normalized relative to the rate that prevails at banks of the same size in the same region

12. Results • Growth, the initial profit drop, and subsequent portfolio management each appear to be important determinants • Baseline recover rate is 0.36 percent • Estimated GDP coefficient suggests that moving from the 25th perc. to the 75th perc. in growth raises probability of recovery by 11.3 percentage points The level of ROA at t=0 has a positive significant coefficient  A bank that underperformed its benchmark by 109 b.p. (25th perc.) compared to one that underperformed by 46 b.p. (75th perc.) is 7 percentage points less likely to recover • The default rate (t+1 to t+3 deviation from region/size mean) is negative and statistically significant A bank that keeps default rate 1.15 percentage points below benchmark (25th perc.) would be 6.9 percentage points more likely to recover than one with average default rate 1.60 percentage points above benchmark (75th perc.)

13. Matching banks and borrowers Merge firms’ balance sheet and income statement data contained in the Company Accounts Data Set (CADS) with data on loans in Italian Central Credit Register (CR): • CADS: proprietary data base containing financial data on a sample of around 25,000 Italian firms (around 49 percent of total sales of nonfinancial firms in the national income accounts) • CADS contains a z score measuring the probability of default on a loan computed with linear discriminant analysis (9 point scale) • CR contains information on loans for all relationships above a threshold of 75,000 euros in loans or commitments (more than 900,000 borrowing firms)

14. Data • Identify all loans in CR made to CADS firms • Select any firm that borrowed at least once from the sample banks between 1986-2001 • Shift to event time, find all borrowers at t=0 & track relationships back to t-1 and forward up to t+3 62,626 credit relationships Focus on the ZSCORE: probability of default & groups firms into 9 categories increasing in risk mapped by CADS into: safe (1, 2), solvent (3, 4), vulnerable (5, 6), risky (7,8,9)

15. Key Variables Two primary measures of credit risk: level of z-score (ZSCORE) and share of all credit in a given bank’s portfolio that is tied up in loans to high risk borrowers (Share HIGH Z). Average values of ZSCORE are 5.15 and 5.25 for the R and NR banks at t=0 Given the # of loans in sample the difference in means is statistically significant: for every 100 borrowers, the distribution of the z’s for the NR banks would have 10 borrowers with a rating of one grade worse than for R banks. The share of high risk loans at t=0 is 20.5% and 21.5% for the R and NR banks: for every 100 euros of loans made, the NR banks are lending one more euro to the high risk borrowers Seems like a small difference

16. Regressions LOANSDOWN=1 if credit granted t=3 is less than t=0, 0 otherwise Prob(LOANSDOWN=1) = f(RECOVER, ZSCORE, Share HIGHZ, FIRM CONTROLS, region dummies, year dummies) + nonlinear effects of key variables Limitation: not entire portfolio; analysis conditional on borrowers affiliated at t=0 But exercise is useful for two reasons: • Corporate lending is largest component of portfolios (loans to households 15% in 1995) and are mainly mortgages (less flexible); • Loans to CADS firms is on average more than 30% of loan portfolio for banks in sample.  These customers are more insulated from credit reductions than others; any effects that we do find understate what might occur for the smaller more typical bank customers

18. Main Results • R banks are 14.4 percentage points more likely to continue extending credit to clients than NR banks • BOTH R & NR banks discriminate between high and lower risk borrower: as z-score moves above 3, more likely to cut credit  Comparing firms with z= 3 & z= 9, probability of reducing credit increases by 40 percentage points • R & NR banks show no significant differences in how they treat borrowers with the same z-scores • Firms that have longer borrowing relationships are more likely to see their credit sustained

19. Main Results Adding Share HIGH Z and interaction with RECOVER: • Coefficient on RECOVER jumps: R banks more likely to continue granting credit to average client once we account for differences based on share of risky borrowers • NR banks with many high risk borrowers are more prone to keep extending credit • R banks become more likely to trim lending when portfolios filled with high risk borrowers. • Difference between NR & R banks is substantial  2 standard deviation change in SHHIGHZ (0.15) implies: • NR banks are 9 percentage points less likely to reduce credit to the typical customer • R banks are 9 percentage points more likely to reduce credit

20. Preliminary Conclusions • Little is known about what governs recovery from banking crises: data uncover several robust patterns  • Banks that get into trouble have been lending to riskier clients than the average in the overall economy (higher interest margin before shock, higher default rate after) • The size of the initial drop in profits is one important factor governing recovery • General business climate after the shock also matters. • Recovery depends on ability to adjust loan portfolio  recovering banks show consistently lower default rates on loans in post-shock period • Micro evidence suggests that recovering banks with large share of high risk customers are more aggressive in trimming loans, while non-recovering banks become less likely to cut credit when portfolio has many high risk loans

21. Extensions • Study new relationships after t=0 • Explore subsamples of banks • Tie results to theory (monitoring, liquidity provisions)