Dynamic Interpretation of Emerging Risks in the Financial Sector

Transcription

1 Dynamic Interpretation of Emerging Risks in the Financial Sector PRESENTER Kathleen Weiss Hanley, Lehigh University Joint work with Gerard Hoberg, University of Southern California

2 National Science Foundation Project made feasible by grant # funded through NSF CIFRAM program. Understanding the economic channels of system-wide risk build-up is important in heading off future crises

3 Existing measures of systemic risk Bisias, Flood, Lo and Valavanis (212) summarize over 3 quantitative systemic risk metrics Liquidity mismatch (Brunnermeier, Gorton and Krishnamurthy, 214), interconnectedness (Billio, Getmansky, Lo and Pelizzon, 212), and bank risk (Adrian and Brunnermeier, 216) to name only a few Quantitative metrics, although useful, have the following drawbacks: General measures: Difficult to identify underlying source of risk Specific measures: Requires a specific theory and may not be useful if source of risk is unknown Using computational linguistics and big data, we crowd source aggregate risks across entire banking industry and present a dynamic measure that is specific about channels

4 Our findings Our method can provide an early warning signal of potential financial instability, identify economic causes and determine which banks may be most affected Aggregate risk score becomes highly significant in 2Q25 well in advance of the financial crisis Economic factors known to contribute to the financial crisis are elevated in the period leading up to Lehman s failure More importantly, we see significant increase in risk build-up in the current period Individual bank exposure to risk themes predicts crises returns, failure and volatility

5 Information production Our methodology requires that both banks and investors produce information Banks Banks are required by SEC to disclose exposure to risks in the 1-K are high-level discussions Useful to investors to determine whether the banking sector has become more risky thereby necessitating additional information production Investors Produce and aggregate information that is manifest in stock returns (Hayek (1945), Grossman and Stiglitz (198) Use covariance of asset returns to measure commonality of risk exposure between banks

6 Emerging risks Propose two methods to detect emerging risks Static model Risks identified from manual inspection of textual data Economic risks that affect the banking sector regardless of time period studied Dynamic model Automated identification of risks Allows different emerging risks to bubble up in each year

7 Corpus of 1-K Bank Risk Factors

8 Latent Dirichlet Allocation (LDA) LDA proposed by Blei, Ng, Jordan, Michael (23) in Journal of Machine Learning Research Proposes that writer is like a hidden Markov Chain who chooses among topics to discuss and then draws words from topic distribution Use Gibbs Sampling to get most likely topics. Goal is to use context to identify interpretable content LDA is automated, replicable and cannot be influenced by researcher bias Our only input is number of topics (25) to be generated

9 LDA topics

10 MetaHeuristica Data

11 Interpretable topic

12 Less interpretable topic

13 LDA limitations Not always interpretable Time-series variation in topics makes comparison difficult Use Semantic Vector Analysis in second stage See Mikolov, Chen, Corrado, and Dean (213) and Mikolov, Sutskever, Chen, Corrado, and Dean (213) Distributional semantics: word is characterized by the company it keeps Firth (1957) Position of word matters

14 Semantic Vector Analysis (SVA) Two stages 1 All 1-Ks are loaded and distributional information about proximity of each word to other words is determined Uses a two layer neural network to Predict a single word given its immediate surrounding words Predict words surrounding a single word 2 Input any word or commongram and the application returns a vector of words with weights indicating importance that best describe that token

15 Semantic theme content Real Estate Deposits Cosine Cosine Row Word Dist Word Dist 1 real.7875 deposits 1 2 estate.7875 deposit foreclosure.4898 brokered deposits property.4619 cdars personal.4563 account registry physical possession.4539 brokered certificates foreclosed real.453 bearing checking foreclosed.4423 bearing deposits deed.4323 certificates beneficiary.4283 negotiable order real estate.4262 promontory interfinancial possession.4147 cdars program oreo.463 sweep ics lien.444 brokered securing.439 withdrawal h2c.414 overdrafts owned.3996 sweep accounts repossessed.3981 bearing death.3974 cdars network owner.3949 fdic insured.455

16 Mapping semantic themes to bank-years

17 Emerging risk model Covariance i,j,t = α + γx i,j,t + ε i,j,t, (1) Covariance i,j,t = α +β 1 S i,j,t,1 +β 2 S i,j,t,2 +β 3 S i,j,t, β T S i,j,t,31 Aggregate risk score +γx i,j,t + ε i,j,t, (2) Take difference in R 2 from Eq. (1) and (2) Scale differential R 2 using its mean and standard deviation from baseline period to get t-statistic in each quarter Elevated t-statistic indicates importance of risk themes and hence, emerging risk

18 Data sources CRSP (stock returns), Compustat (accounting variables) FDIC Failures and Assistance Transactions List VIX data. Call Reports for bank-specific characteristics metaheuristica used to extract risk factor discussions from bank 1-Ks from 1997 to 214 Include banks defined as having SIC codes from 6 to 6199 Require machine readable 1-K, with some non-empty discussion of risk factors

19 Static risk method

20 Determining static themes Examine LDA output and feed prevalent (most frequent) key phrases (tokens) from LDA to SVA These are high-level risk factors that remain constant over time Remove any boilerplate such as balance sheet or million December Group the remaining individual terms into broad categories of risks fundamental to the banking sector aided by a review of the literature e.g. Credit Card or Regulatory Capital For our static model, we choose 61 initial semantic themes upon reviewing the LDA output for key phrases and reduce this to 31 themes due to multicollinearity

21 Static semantic themes Accounting Cash Certificate Deposit Commercial Paper Compensation Competition Counterparty Credit Card Currency Exchange Data Security Deposits Derivative Dividends Fees Funding Sources Governance Growth Strategy Insurance Internal Controls Lawsuit Mergers Acquisitions Off Balance Sheet Operational Risk Prepayment Rating Agency Real Estate Regulatory Capital Reputation Securitization Student Loans Taxes

22 Aggregate risk metric Run regression once per quarter with one observation bank-pair (i and j). Dependent variable is quarterly return covariance of bank i and j measured using daily returns Semantic theme of pair is the product S i,j = S i S j X is a set of pairwise controls including size, age, profitability, leverage, and industry controls Aggregate risk score is the contribution of SVA themes to R 2

23 Aggregate emerging risk score z score

24 Other emerging risk metrics VIX Level Std Dev Returns (Financials) EPU USA

25 Identifying individual risks Use each of 31 semantic themes from SVA We compute the individual contribution to R 2 of each theme in explaining pairwise return covariance in each quarter Standardize each marginal R 2 by its mean and standard deviation from the baseline period 1998 to 23 Resulting t-statistics illustrate how strong each individual risk factor is in explaining comovement Importantly, individual risk factors are interpretable This has important ramifications both for understanding the crisis and monitoring emerging risk in the current period.

26 28 major risks Real Estate Prepayment Commercial Paper Credit Card Dividend Rating Agency Operational Risk

27 215 major risks Mergers Acquisition Cash Real Estate Lawsuit Taxes Counterparty Operational Risk

28 Drill-down model: Real estate Subprime Mortgage Backed Freddie Mac &Fannie Mae Heloc Foreclosed

29 Dynamic methodology Extract top 25 terms from each of the 25 LDA topics per year (625 possible topics per year) Limit to bigrams (4 possible topics per year) Remove boilerplate (15 possible topics per year) Use covariance model and stepwise regression to maximize R 2 Baseline R 2 measured using four year moving window of adjusted R 2 ending in the year being tested

30 Dynamic emerging risks Emerging Risk Year Emerging Risk Year related litigation 241 economic downturn 2113 deposits borrowings 241 education loans 2113 mortgage banking 243 identity theft 2113 operational risk 243 customer deposits 2114 charged off 243 secondary mortgage 2121 origination fees 244 deposit insurance 2122 backed securities 244 foreclosure process 2122 off balance 252 commercial real 2123 rate environment 252 operational risk 2124 real estate 253 trust preferred 2132 rate swap 254 extend credit 2132 recruiting hiring 261 weather events 2133 board directors 262 executive compensation 2133 interest bearing 262 supervision regulation 2134 underwriting standards 263 regulatory requirements 2134 time deposits 264 basel iii 2141 brokered deposits 264 negative publicity 2142 investment securities 264 supervision regulation 2142 senior notes 271 capital levels 2143 board directors 272 regulatory authorities 2143 prevent fraud 273 brokered deposits 2144 damage reputation 274 senior management 2151 extend credit 274 legal proceedings 2161 cost funds 281 servicing rights 2161 rate risk 282 institution failures 2161 real property 283 merger agreement 2163 legal proceedings 284 credit risk 2163 mergers acquisitions 291 data processing 2164

31 Individual bank exposure to emerging risk Create Emerging Risk Exposure as average quarterly predicted covariance bank i has with all other banks j using the main covariance model in Equation (2) Uses the following procedure: 1 Take product of fitted coefficients for each SVA theme (β 1 to β 31 ) from the baseline covariance model and multiply by the given bank-pair s SVA theme loading 2 Sum the resulting 31 products for each bank-pair to get the total predicted covariance of bank i with each bank j 3 Average predicted covariances over banks j to get the total Emerging Risk Exposure for bank i in quarter t

32 Cross-sectional tests using static model In each quarter, run single cross sectional regression Dependent variable is one of the following: Bank s stock return from 9/28 to 12/212 Bank s stock return from 12/215 to 2/216 Dummy variable indicating whether the given bank failed in the 3 year period beginning with the Lehman bankruptcy Also run monthly Fama-McBeth regressions where dependent variable is the ex post monthly stock return volatility computed using daily stock returns. Main independent variable of interest is Emerging Risk Exposure

33 Predicting post-28 crisis returns (9/28-12/212) Emerging Risk # Predictive Row Quarter Exposure Obs Timing (1) 24 1Q 2.41 (2.16) 352 Predictive (2) 24 2Q (3.69) 352 Predictive (3) 24 3Q.319 (.18) 368 Predictive (4) 24 4Q.415 (.28) 368 Predictive (5) 25 1Q -.67 (-.31) 388 Predictive (6) 25 2Q (-.28) 388 Predictive (7) 25 3Q -1.6 (-.36) 418 Predictive (8) 25 4Q (.4) 418 Predictive (9) 26 1Q.918 (.65) 47 Predictive (1) 26 2Q (-1.44) 47 Predictive (11) 26 3Q (-1.6) 43 Predictive (12) 26 4Q (-1.9) 43 Predictive (13) 27 1Q (-2.1) 444 Predictive (14) 27 2Q (-2.1) 444 Predictive (15) 27 3Q (-3.4) 469 Predictive (16) 27 4Q (-3.27) 469 Predictive (17) 28 1Q (-3.65) 468 Predictive (18) 28 2Q (-7.8) 468 Predictive (19) 28 3Q (-2.21) 489 Non Predictive (2) 28 4Q (-1.2) 491 Non Predictive (21) 29 1Q (-1.15) 518 Non Predictive (22) 29 2Q (-1.55) 518 Non Predictive (23) 29 3Q (-9.97) 529 Non Predictive (24) 29 4Q (-2.88) 522 Non Predictive

34 Predicting current period returns (12/215-2/216) Emerging Risk # Predictive Row Quarter Exposure Obs Timing (1) 21 1Q (-3.25) 334 Predictive (2) 21 2Q (-3.27) 334 Predictive (3) 21 3Q (-4.44) 341 Predictive (4) 21 4Q (-1.53) 341 Predictive (5) 211 1Q (-3.33) 351 Predictive (6) 211 2Q (-4.22) 35 Predictive (7) 211 3Q (-11.7) 356 Predictive (8) 211 4Q (-4.3) 356 Predictive (9) 212 1Q (-2.4) 349 Predictive (1) 212 2Q -.66 (-1.4) 349 Predictive (11) 212 3Q (-3.73) 36 Predictive (12) 212 4Q (-1.77) 36 Predictive (13) 213 1Q (-1.45) 351 Predictive (14) 213 2Q (-1.92) 351 Predictive (15) 213 3Q.198 (.95) 368 Predictive (16) 213 4Q (-2.54) 368 Predictive (17) 214 1Q -.24 (-.17) 356 Predictive (18) 214 2Q (-3.) 356 Predictive (19) 214 3Q (-2.42) 367 Predictive (2) 214 4Q (-2.3) 367 Predictive (21) 215 1Q -.44 (-1.53) 358 Predictive (22) 215 2Q -.55 (-1.47) 358 Predictive (23) 215 3Q (-2.33) 387 Predictive (24) 215 4Q -.5 (-1.49) 386 Non Predictive

35 Predicting bank failures Emerging Risk Predictive Quarter Exposure s Obs Timing 24 1Q.4 (.8) 625 Predictive 24 2Q.4 (.94) 625 Predictive 24 3Q -.5 (-1.3) 625 Predictive 24 4Q -.4 (-.79) 625 Predictive 25 1Q -.2 (-1.33) 615 Predictive 25 2Q -.1 (-1.36) 615 Predictive 25 3Q.8 (3.56) 615 Predictive 25 4Q.6 (2.55) 615 Predictive 26 1Q -.2 (-.14) 578 Predictive 26 2Q -.1 (-.8) 578 Predictive 26 3Q.3 (.58) 578 Predictive 26 4Q.8 (3.97) 578 Predictive 27 1Q.9 (3.96) 588 Predictive 27 2Q.11 (7.36) 588 Predictive 27 3Q.1 (2.31) 588 Predictive 27 4Q.14 (4.37) 588 Predictive 28 1Q.14 (4.42) 562 Predictive 28 2Q.15 (3.89) 562 Predictive 28 3Q.15 (3.72) 562 Predictive 28 4Q.4 (.63) 562 Non Predictive 29 1Q.24 (8.54) 564 Non Predictive 29 2Q.1 (3.87) 564 Non Predictive 29 3Q -.1 (-.27) 564 Non Predictive 29 4Q.7 (1.96) 564 Non Predictive

36 Unconditional Fama-MacBeth volatility regressions 1 Quarter 2 Quarter 3 Quarter Lag Exposure Exposure Exposure Obs (8.94).15 (1.26).112 (11.35) (8.72).14 (1.22).18 (11.13) (9.18).99 (1.53).14 (11.38) (9.13).98 (1.81).12 (11.43) (9.13).93 (1.42).97 (11.32) (8.96).88 (1.4).88 (11.9) (9.52).83 (1.66).81 (1.52) (8.66).77 (1.4).74 (9.6) (8.59).69 (9.39).71 (9.9) (8.65).64 (8.62).66 (8.82) (8.38).6 (8.28).63 (8.51) (7.51).57 (7.74).6 (8.6) (6.84).49 (7.4).54 (7.43) (6.29).46 (6.79).51 (6.95) (5.81).44 (6.49).47 (6.56) (5.9).4 (5.54).43 (5.83) (4.63).4 (5.4).42 (5.61) (4.73).39 (5.25).42 (5.6) (4.2).36 (4.73).41 (5.27) (4.62).36 (5.).41 (5.3) (4.26).35 (4.99).39 (5.12) (4.16).36 (5.24).39 (5.25) (3.8).34 (4.68).36 (4.86) (4.23).34 (4.59).35 (4.72) (4.24).35 (4.34).35 (4.5) (3.6).31 (3.8).33 (4.14) (3.43).29 (3.65).33 (4.1) (3.36).3 (3.85).33 (4.2) (3.17).3 (3.95).34 (4.46) (2.65).27 (3.53).29 (3.78) (2.61).24 (3.19).26 (3.46) Hanley.22 and (3.8) Hoberg (218).26 (3.54) Jacobs.25 Levy (3.32) Equity Management Center Conference

37 Conclusions We propose a model of emerging risks in the financial sector based on computational linguistic analysis of firm disclosures and return covariances Method is flexible, dynamic, timely, allowing the prediction of interpretable emerging risks for which a researcher might not even be aware Allows for high-level (aggregate) to granular level (theme and bank) determination of risk build-up Can be used by researchers and regulators alike to monitor threats to financial stability