## wp17180

## Source details

**Canonical URL:** [wp17180](https://www.imf.org/-/media/files/publications/wp/2017/wp17180.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2017/wp17180.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2017/wp17180.pdf.json)

---

### 3.1 Insurers — Data, measurement and identification
- Data and coverage:
  - Data source: detailed insurance regulatory data from SNL Financial (2005-2014) at position and transaction level for bonds in the general account and for derivatives (transaction level).
  - Coverage: detailed assets in general account comprise around half of insurers’ total assets and roughly correspond with life insurance liabilities; within the general account, most investment is in bonds.
  - Sample construction: restrict to large insurers (top two asset deciles by quarter): 64 insurers that account for 90% of life insurers’ general account bond holdings; 50 insurers matched to corporate parent; quarterly sample 1,701 insurer-quarters.
  - Frequency and instruments: quarterly holdings at CUSIP-level (Schedule D) and interest-rate derivatives outstanding at transaction level (Schedule DB); daily view of transactions includes transaction-driven changes (excludes mark-to-market; mark-to-market appears in quarterly view).
- Risk measures and pricing:
  - Credit-risk measure: average yields to maturity on bonds purchased by insurers in a given quarter (transaction-level purchases), following Becker & Ivashina (2015).
  - Interest-rate risk measure: DV01 (absolute change in value when interest rates increase by one basis point, parallel shift). Bonds held tend to have duration around 8 years.
  - Derivatives pricing: each contract priced using a log-linearly interpolated swap curve (Bloomberg); swap DV01 calculated numerically.
- Liability-side approximation and net exposure:
  - Liability-side data limitation: only market value of liabilities available. Assumed average modified duration of liabilities = 15 years.
  - Example modified duration by yield (5% coupon average values shown): 13.1 (6%), 15.4 (4%), 18.1 (2%).
  - Net DV01 Gap defined as (liability-side DV01 − sum(bond DV01 + derivatives DV01)) / liability-side DV01, expressed in percentage points (positive when assets underweight liability duration). Bond DV01 gap similar but excludes derivatives.
  - DV01 gaps are roughly 40% on average.
- Swap portfolios and scale:
  - Life insurers hold large offsetting pay-fixed and receive-fixed interest-rate swap positions; net positions typically receive-fixed (adds asset-side duration).
  - Illustrative quarter (2012 Q2): total swap notional ≈ $600BN; net receive-fixed notional < $100BN; duration-equivalent notional ≈ $200BN; general account bond holdings ≈ $2TN.
  - For the US life insurer sample, derivatives account for about 10% of the exposure of life insurers’ general account bonds.
- Bond transaction and new-issue matching:
  - New-issue analysis combines Dealogic new-issue data (2005-2014) with daily bond-level transaction data via CUSIP; Dealogic issuer classification used to identify issuers.
- Empirical classification and identification:
  - Difference-in-difference framework comparing large financial institutions “hit hard” by the crisis to those “hit less hard.”
  - Insurer “hit hard” flag (corporate parent level; mutuals excluded from baseline): flag if, during 2008-2010, dividend growth or changes in book equity-to-asset ratios fall below the 10th percentile at any point, or if any equity was issued. This flags 24 out of the 50 matched insurers as “hit hard.”
  - Time periods: Pre-crisis 2005Q1-2007Q2; Crisis 2007Q3-2010Q4; Post-crisis 2011Q1-2014Q4.
  - Baseline dependent variables: Net DV01 Gap (percentage points) and average yield to maturity of bonds purchased (basis points). Specifications include insurer-level crisis-hit dummy, interactions with crisis and post-crisis periods, controls for size, quarter fixed effects; some specifications include both quarter and insurer fixed effects. Standard errors double-clustered by quarter and insurer.

### 3.1 Insurers — Main empirical findings
- Directional result:
  - Large life insurers hit hard by the crisis reduced risk taking relative to large insurers hit less hard.
- Interest-rate risk:
  - Insurers hit hard reduced their net DV01 gap by 8 percentage points (four-tenths of a standard deviation) during and after the crisis relative to insurers hit less hard.
  - Results hold with both quarter and insurer fixed effects.
  - The reduction in net DV01 gap is not explained by simple purchases of lower-duration bonds; insurers hit hard bought bonds with higher duration in some specifications.
- Credit risk:
  - Insurers hit hard purchased bonds yielding on average 30 basis points less than insurers hit less hard (within-quarter standard deviation of this variable ranges from 30-120 basis points over the sample).
- Timing and dynamics:
  - Pre-crisis both groups moved in parallel. Divergence emerges during crisis and post-crisis with hit-hard insurers pulling back. Panel evidence shows convergence on DV01 gaps and persistent gap in yields of bonds bought starting at beginning of crisis through 2012.
- Interpretation:
  - Results run counter to a simple “risk-shifting” or “gamble-for-resurrection” prediction.
  - Findings are consistent with insurers valuing franchise value and avoiding actions that increase the likelihood or cost of failure.

### 3.1 Insurers — Additional methodological notes (selected)
- Bond pricing and DV01:
  - Yield-to-maturity solved numerically using semi-annual payment assumption (f = 2) with a 20-iteration bisection search; reported fair value must be within 1% of estimated fair value or observation dropped.
  - DV01 calculated numerically as price change for a 1% shift in yield to maturity.
- Derivatives: interest-rate swaps
  - Coverage 2005-2014 quarterly; reporting schedules differ pre-2010 and post-2010 but swaps consistently available.
  - Swap direction and fixed rate parsed from text; hierarchy of identification includes checks against fair value signs and interpolated swap curve; DV01 calculated as change in fair value given a one basis point shift in the swap curve.

---

### 4.2 The role of regulation — Question and approach
- Key question:
  - Whether regulation (formal rules or informal regulator actions) can explain why insurers "hit hard" pulled back risk after the crisis, as an alternative to franchise value explanations.
- Empirical leverage:
  - Insurance RBC requirements are applied consistently across insurers and were not raised for insurers hit hard by the crisis.
  - Insurance regulatory data allow analysis of variation within regulatory buckets (coarse, rating-based NAIC categories) that receive identical regulatory treatment.
  - Empirical strategy uses daily transaction-level data for newly issued bonds close to issuance, with more than 85,000 individual transactions.

### 4.2 The role of regulation — Evidence against a pure regulation explanation
- Within-bucket risk reduction:
  - Sample: investment grade corporate bonds (excluding financial issuers) issued 2005-2014 with at least 10 transactions by large insurers with matched parents within three months of first insurer transaction.
  - Relative to pre-crisis, insurers hit hard accounted for lower fractions of purchases of newly issued bonds during and after the crisis.
  - The effect of yields on relative purchasing by insurers hit hard is strongly statistically significant during and after the crisis.
  - Insurers hit hard reduced risk taking within NAIC categories (NAIC 1 and NAIC 2), i.e., within bonds that face identical RBC treatment.
  - For NAIC 1 corporate bonds (AAA to A), relative to pre-crisis, during and after the crisis the share of purchases by insurers hit hard was 22-30 percentage points lower for a bond with a one percentage point higher yield.
- Robustness:
  - Results hold when broadening sample to include financial issuers and private structured bonds.
  - Results consistent with Becker & Ivashina (2015) and may explain part of their finding because insurers hit hard tended to be larger.
- Interpretation:
  - Risk reduction within assets with identical regulatory treatment implies capital requirements alone cannot explain observed risk reduction by insurers hit hard.
  - Supports explanations based on franchise value or other financial-frictions-based mechanisms that induce risk reduction without changing explicit regulatory capital constraints.

### 4.2 The role of regulation — Informal regulatory actions and moral suasion
- Considerations limiting moral suasion explanation:
  - US insurance regulation is fragmented at state level: more than fifty separate insurance commissioners; assets not highly concentrated by state.
  - The top 10 states cover three quarters of general account assets for large life insurers with matched parents; effective moral suasion across states would require close coordination, which is unlikely.
  - Financial Analysis Working Group’s role limited to applying "peer pressure" on the lead regulator (NAIC 2013).
  - Documented cross-state variation in enforcement and opportunities for cross-state regulatory arbitrage.
  - Capital requirements for MBS were lowered, not raised, during the crisis, making a generalized regulatory push against MBS inconsistent with observed capital-rule changes.
- Synthesis:
  - Financial frictions that push toward risk reduction after negative shocks go beyond explicit regulatory constraints.
  - Evidence of within-bucket risk reduction supports franchise value or financial-frictions mechanisms.
  - While informal regulatory pressure cannot be definitively ruled out, fragmented state-level structure makes coordinated moral suasion an unlikely sole explanation.

---

### 4.4 Results for banks — Measures, findings and implications
- Measures and empirical setup:
  - Baseline credit-risk measure for banks: natural log changes in real estate loans (BHC data), multiplied by 100.
  - Baseline interest-rate-risk measure for banks: fraction of loans with maturity greater than one year (call report data), in percentage points.
  - Regression structure parallels insurers’ analysis: Crisis Hit Flag at parent level, interactions with crisis and post-crisis dummies, Log(Assets), quarter fixed effects; standard errors double-clustered by bank and quarter.
  - Identification: subset of large banks identified as “hit hard by the crisis” based on corporate parent information.
- Main empirical findings:
  - Large banks hit hard by the crisis pulled back from risk taking relative to banks not hit as hard.
  - Credit portfolios of banks hit hard grew slower than those of banks hit less hard.
  - During the crisis, real estate loan portfolios grew by one percentage point per quarter less for banks hit hard (a third of a standard deviation).
  - Post-crisis, banks hit hard increased their loan maturities less; almost all maturity extension concentrated within banks not hit as hard.
  - Both groups moved in parallel pre-crisis (parallel trends).
- Robustness:
  - Results apply to total lending as well as real estate lending (Table 10), indicating results are not driven by changes in loan loss provisions because total loans are reported net of loan loss provisions while components are not.
  - Parallel trends in loan growth extend back to 2000.
  - TARP issuance and repayment did not differentially affect risk taking by banks hit hard.
- Comparisons with insurers and interpretation:
  - Both insurers and banks hit hard pulled back from risk taking, but sectoral differences in liability structure imply different margin of adjustment:
    - Insurers reduced interest-rate risk by increasing asset duration, reducing the gap to liability duration.
    - Banks reduced interest-rate risk by not lengthening asset maturities (they take interest-rate risk by lengthening asset maturities while borrowing short).
  - Opposite directional shifts across sectors indicate neither sector took directional bets on interest rates.
- Magnitude and broader implications:
  - The difference in real estate loan growth between banks hit hard and banks hit less hard can account for about half of the total reduction in real estate lending during the crisis.
  - Policy implications:
    - If franchise value deters risk taking in some circumstances, macro‑prudential policy aiming to loosen regulatory constraints should focus on constraints that actually bind when institutions are in trouble.
    - Time varying capital requirements may not be sufficient to improve credit supply if reduced risk taking is driven by forces other than capital requirements.
    - Actions that make financial institutions’ future clear—such as stress tests, recapitalization, and clear resolution frameworks—might be more effective.

---

### Selected quantitative highlights (preserved exactly as in source)
- Sample sizes and scope:
  - Insurer-Quarters: 1,701; Insurers with matched parents: 50; Insurers total in some tables: 64.
  - BHC-Quarters: 2,144; BHCs: 54.
  - Issues in Table 3: 3,058; Transactions: 66,056. Issues in Table 4: 4,144; Transactions: 86,778.
- Key coefficients and estimates (selected, preserved exactly):
  - Insurers: Insurers hit hard reduced net DV01 gap by 8 percentage points (four-tenths of a standard deviation).
  - Insurers: Insurers hit hard purchased bonds yielding on average 30 basis points less than insurers hit less hard.
  - Table 3 (All IG): YTM coefficient 12.0 (t-statistic 4.17); YTM×2007Q3-2010Q4: -14.9 (t-statistic -5.31); YTM×2011Q1-2014Q4: -17.7 (t-statistic -5.91).
  - Table 5 (Including derivatives): Crisis Hit Flag×2011Q1-2014Q4: -8.96 (t-statistic -2.74); R 2: 0.09; Insurer-Quarters: 1,701; Insurers: 50.
  - Table 6 (YTM): Crisis Hit Flag×2007Q3-2010Q4: -28.28 (t-statistic -2.45); Log(Assets): -19.26 (t-statistic -2.85); R 2 (YTM): 0.63.
  - Banks (Table 9, RE loan growth): Crisis Hit Flag×2007Q3-2010Q4: -1.02 (t-statistic -2.47); R 2 (RE loan growth): 0.22.
  - Descriptive (Table 1): Assets (BN) 92.3 38.6 135.8 113.7; Equity/Assets 8.0 9.6 6.8 4.8; Net DV01 Gap 40.3 39.0 41.4 20.9; YTM (Purchases) 470.9 484.7 459.7 112.7; Dur (Purchases) 10.0 9.6 10.3 2.1; Insurer-Quarters 1,701; Insurers 50.
  - Bank descriptive (Table 8): Assets 184.0 / 148.3 / 226.0 / 473.6; Loan Growth 1.1 / 1.2 / 1.0 / 2.7; RE Loan Growth 0.7 / 1.0 / 0.4 / 2.8; Loans>1 Year 47.1 / 48.2 / 45.7 / 17.2.

*Source: wp17180 (IMF Working Paper chapter).*

### 3.1  Insurers

### 3.1  Insurers

### Data and measurement
- Data source: detailed insurance regulatory data from SNL Financial (2005-2014) at position and transaction level for bonds in the general account and for derivatives (transaction level).  
- Coverage: detailed assets in general account comprise around half of insurers’ total assets and roughly correspond with life insurance liabilities. Within the general account, most investment is in bonds.  
- Sample construction:
  - Restrict to large insurers (top two asset deciles by quarter): 64 insurers that account for 90% of life insurers’ general account bond holdings.
  - Matched to corporate parent level for 50 insurers.
  - Quarterly sample: 1,701 insurer-quarters.  
- Frequency and instruments:
  - Quarterly holdings at CUSIP-level (Schedule D) and interest-rate derivatives outstanding at transaction level (Schedule DB).
  - Daily view of transactions includes transaction-driven changes (excludes mark-to-market; mark-to-market appears in quarterly view).
- Credit-risk measure: average yields to maturity on bonds purchased by insurers in a given quarter (transaction-level purchases), following Becker & Ivashina (2015).
- Interest-rate risk measure: DV01 (absolute change in value when interest rates increase by one basis point, parallel shift). Bonds held tend to have duration around 8 years. To price derivatives, each contract is priced using a log-linearly interpolated swap curve (Bloomberg).

### Liability-side approximation and net exposure
- Liability-side data limitation: only market value of liabilities available. Assumption used: average modified duration of liabilities = 15 years.
- Example duration table (modified duration for bonds with 5% coupon; average values shown):
  - Average modified duration scenarios by yield: 13.1 (6%), 15.4 (4%), 18.1 (2%).
- Net duration exposure:
  - Net DV01 Gap defined as (liability-side DV01 − sum(bond DV01 + derivatives DV01)) / liability-side DV01, expressed in percentage points (positive when assets underweight liability duration).  
  - Bond DV01 gap similar but excludes derivatives.  
  - DV01 gaps are roughly 40% on average.

### Swap portfolios and scale of derivatives exposure
- Life insurers hold large offsetting pay-fixed and receive-fixed interest-rate swap positions; net positions typically receive-fixed (which adds asset-side duration).
- Illustrative quarter (2012 Q2):
  - Total swap notional ≈ $600BN; net receive-fixed notional < $100BN.
  - Duration-equivalent notional ≈ $200BN.
  - General account bond holdings valued at ≈ $2TN.
- Comparative markets and scale:
  - BIS (first half 2016): interest-rate derivatives notional $420TN, gross market value $15TN; total market value of corporate and foreign bonds $12TN.
  - For the US life insurer sample, derivatives account for about 10% of the exposure of life insurers’ general account bonds.
- Derivative pricing: individual contracts priced to produce a measure of interest-rate exposure consistent across bonds and derivatives.

### Bond-level analysis and new issue matching
- New-issue analysis: combine Dealogic new-issue data (2005-2014) with daily bond-level transaction data via CUSIP; Dealogic issuer classification used to identify issuers.
- Purpose: to analyze risk taking within regulatory categories and within newly issued bonds.

### Empirical classification and identification strategy
- Difference-in-difference framework comparing large financial institutions “hit hard” by the crisis to those “hit less hard.”
- Insurer “hit hard” flag (corporate parent level; mutuals excluded from baseline):
  - Flag if, during 2008-2010, dividend growth or changes in book equity-to-asset ratios fall below the 10th percentile at any point, or if any equity was issued.
  - This flags 24 out of the 50 large insurers with matched parents as “hit hard” (examples: AIG, Prudential, Metlife flagged; Berkshire Hathaway not flagged).
- Time periods:
  - Pre-crisis: 2005Q1-2007Q2
  - Crisis: 2007Q3-2010Q4
  - Post-crisis: 2011Q1-2014Q4
- Regression controls and inference:
  - Baseline dependent variables: Net DV01 Gap (percentage points) and average yield to maturity of bonds purchased (basis points).
  - Specifications include insurer-level crisis-hit dummy, interactions with crisis and post-crisis periods, controls for size, quarter fixed effects; some specifications include both quarter and insurer fixed effects.
  - Standard errors double-clustered by quarter and insurer.

### Main empirical findings for insurers
- Directional result: large life insurers hit hard by the crisis reduced risk taking relative to large insurers hit less hard.
- Interest-rate risk:
  - Insurers hit hard reduced their net DV01 gap by 8 percentage points (four-tenths of a standard deviation) during and after the crisis relative to insurers hit less hard.
  - Results hold with both quarter and insurer fixed effects.
  - The reduction in net DV01 gap is not explained by simple purchases of lower-duration bonds; if anything, insurers hit hard bought bonds with higher duration (discussed in subsequent sections).
- Credit risk:
  - Insurers hit hard purchased bonds yielding on average 30 basis points less than insurers hit less hard (within-quarter standard deviation of this variable ranges from 30-120 basis points over the sample).
- Timing and dynamics:
  - Pre-crisis, both groups moved in parallel. Divergence emerges during crisis and post-crisis with hit-hard insurers pulling back.
  - Figure evidence: Panel A (interest-rate risk) and Panel B (credit risk) show convergence on DV01 gaps and persistent gap in yields of bonds bought starting at beginning of crisis through 2012.
- Interpretation:
  - Results run counter to a simple “risk-shifting” or “gamble-for-resurrection” prediction (which would predict increased risk taking by institutions hit harder).
  - Findings are consistent with insurers valuing franchise value and avoiding actions that increase the likelihood or cost of failure.
  - Alternative interpretations noted in the text: risk shifting may occur in other circumstances (smaller or idiosyncratic shocks), may require even larger shocks, or may be mitigated by contracting arrangements or other incentives (e.g., earnings smoothing).

*Source: wp17180 - 3.1  Insurers (IMF working paper chapter).*

### 4.2  The role of regulation

### 4.2  The role of regulation

### Key question and approach
- Examines whether regulation (formal rules or informal actions by regulators) can explain why insurers "hit hard" pulled back risk after the crisis, as an alternative to franchise value explanations.
- Focuses on insurance because:
  - Risk-based capital (RBC) requirements are applied consistently across insurers and were not raised for insurers hit hard by the crisis.
  - Insurance regulatory data include bond-level holdings and transactions, allowing analysis of variation within regulatory buckets (coarse, rating-based buckets) that receive identical regulatory treatment.
- Empirical strategy uses daily transaction-level data for newly issued bonds close to issuance, with more than 85,000 individual transactions.

### Capital requirements and how they might matter
- Even if capital requirements remained unchanged, insurers hit hard faced larger shocks, which could make capital requirements effectively more binding and create incentives to recapitalize rather than to engage in general risk reduction.
- However, because RBC treatment is coarse and rating-based, variation in risk within identical regulatory buckets can be observed and used to separate regulatory effects from other motives.
- Becker & Ivashina (2015) exploit similar within-bucket variation to show insurers "reach for yield" relative to mutual funds and pension funds; this paper builds on that approach but focuses on within-insurer, bond-level variation.

### Bond-level evidence that limits a pure regulation explanation
- Sample: investment grade corporate bonds (excluding financial issuers) issued from 2005-2014 with at least 10 transactions by large insurers with matched parents within three months of the first insurer transaction.
- Findings:
  - Relative to the pre-crisis period, insurers hit hard accounted for lower fractions of purchases of newly issued bonds during and after the crisis.
  - The effect of yields on the relative purchasing behavior of insurers hit hard is strongly statistically significant both during and after the crisis.
  - Insurers hit hard reduced risk taking within NAIC categories (NAIC 1 and NAIC 2), i.e., within bonds that face identical risk-based capital requirements.
  - For NAIC 1 corporate bonds (AAA to A), relative to the pre-crisis period, during and after the crisis the share of purchases by insurers hit hard was 22-30 percentage points lower (about one to one and a half standard deviations) for a bond with a one percentage point higher yield (less than one standard deviation).
- Robustness:
  - Results hold when broadening the sample to include financial issuers and private structured bonds.
  - Results are consistent with Becker & Ivashina (2015) and may explain part of their finding because insurers hit hard tended to be larger.
- Interpretation:
  - Risk reduction within assets with identical regulatory treatment implies capital requirements alone cannot explain the observed risk reduction by insurers hit hard.
  - This supports explanations based on financial frictions or franchise value constraints that drive risk reduction after negative shocks, and indicates these actions can occur without changing regulatory constraints.

### Informal regulatory actions and moral suasion
- Regulators might have informally pushed insurers hit hard to reduce risk ("moral suasion"), but:
  - US insurance regulation is fragmented at the state level: more than fifty separate insurance commissioners; assets are not highly concentrated by state.
  - The top 10 states cover three quarters of general account assets for large life insurers with matched parents.
  - Effective moral suasion across states would require close coordination among many regulators, which is unlikely. The Financial Analysis Working Group's role is limited to applying "peer pressure" on the lead regulator (NAIC 2013).
  - There is documented cross-state variation in enforcement (e.g., marking to market) and opportunities for cross-state regulatory arbitrage.
- Additional considerations:
  - Moral suasion is more plausible across asset classes (e.g., shifting away from risky MBS), but the paper documents risk reduction at the bond level within investment grade corporate bonds.
  - Capital requirements for MBS were lowered, not raised, during the crisis, making a generalized regulatory push against MBS less consistent with observed capital-rule changes.

### Synthesis
- Financial frictions that push toward risk reduction after negative shocks go beyond explicit regulatory constraints.
- Evidence of risk reduction within identical regulatory buckets supports a role for franchise value or other financial-frictions-based mechanisms that induce insurers hit hard to reduce risk without changing the tightness of regulatory capital constraints.
- While informal regulatory pressure targeted at insurers hit hard cannot be definitively ruled out, the fragmented state-level regulatory structure in the US makes coordinated moral suasion an unlikely sole explanation.
- A broad, across-the-board regulatory push to reduce risk would not have been ideal from a macro-prudential perspective.

*Source: wp17180 - 4.2  The role of regulation*

### 4.4  Results for banks

### 4.4  Results for banks

### Measures and empirical setup
- Baseline credit-risk measure for banks: natural log changes in real estate loans (BHC data), multiplied by 100.
- Baseline interest-rate-risk measure for banks: fraction of loans with maturity greater than one year (call report data), in percentage points.
- Regression structure: same as corresponding analysis for insurers (Table 2). Controls include size directly, quarter fixed effects in all specifications. Standard errors are double clustered.
- Identification: subset of large banks identified as “hit hard by the crisis” based on information about the highest corporate parent (as described in Section 3.3).

### Main empirical findings for banks
- Large banks hit hard by the crisis pulled back from risk taking relative to banks not hit as hard.
- Credit portfolios of banks hit hard grew slower than those of banks hit less hard.
- Banks hit hard did not extend asset maturities post-crisis.
- During the crisis, real estate loan portfolios grew by one percentage point per quarter less for banks hit hard (a third of a standard deviation).
- Post-crisis, banks that were hit hard increased their loan maturities less.
- Within large banks, almost all maturity extension is concentrated within banks that were not hit as hard during the crisis; the difference-in-difference coefficient is more than a third of a standard deviation.
- Both groups of banks moved in parallel pre-crisis (parallel trends).

### Robustness exercises
- Table 10: Results apply to total lending as well as real estate lending. This indicates results are not driven by changes in loan loss provisions because total loans are reported net of loan loss provisions on call reports, while components of lending are reported without deducting loan loss provisions (see Appendix A.2).
- Parallel trends in loan growth shown in Panel B of Figure 7 extend back to 2000.
- TARP issuance and repayment did not differentially affect risk taking by banks hit hard.
- Figures and tables are available on request.

### Comparisons with insurers and interpretation
- Similar approach applied to insurers and banks yields consistent results: institutions hit hard during the crisis pulled back from risk taking.
- Key difference in how interest-rate risk is taken:
  - Insurers: likely take interest-rate risk by not fully offsetting liability duration exposure; insurers hit hard reduced interest-rate risk by increasing asset duration, thereby reducing the gap to their liability duration.
  - Banks: likely take interest-rate risk by lengthening asset maturities while still borrowing short; banks hit hard reduced interest-rate risk by not increasing asset duration.
- These opposite directional shifts across sectors indicate that neither sector was taking directional bets on interest rates.
- Differences in liability structure, regulatory structure, and bailout probabilities do not overturn the core finding that both insurers and banks hit hard pulled back from risk taking.
- The difference in liability structure helps interpret what “reducing risk” means in each sector and supports the argument that large financial institutions reduced risk based on appropriate sector-specific definitions.

### Magnitude and broader implications
- The difference in real estate loan growth between banks hit hard and banks hit less hard can account for about half of the total reduction in real estate lending during the crisis.
- Both insurers and banks hit hard by the crisis reduced risk taking on multiple dimensions.
- The results add to the discussion of why credit supply fell during the crisis by suggesting a role for reduced risk taking by banks affecting credit supply.
- Policy implications:
  - If franchise value deters risk taking in some circumstances, macro‑prudential policy aiming to loosen regulatory constraints should focus on those constraints that actually bind when institutions are in trouble.
  - Time varying capital requirements may not be sufficient to improve credit supply if reduced risk taking is driven by forces other than capital requirements.
  - Actions that make financial institutions’ future clear—such as stress tests, recapitalization, and clear resolution frameworks—might be more effective.

*Source: wp17180 - 4.4  Results for banks*

### References

### wp17180 - References

### References (selected)
- List of cited works includes AAA (1999), AAA (2014), Acharya & Viswanathan (2011), Becker & Ivashina (2015), Becker & Opp (2014), Begenau, Piazzesi & Schneider (2015), Berends & King (2015), Berends et al. (2013), Bidder, Krainer & Shapiro (2017), Brewer III, Mondschean & Strahan (1993), Briys & De Varenne (1997), Brunnermeier & Sannikov (2014), Chodorow-Reich (2014), Chodorow-Reich, Ghent & Haddad (2016), Coimbra & Rey (2017), Cooper et al. (2010), Coval, Jurek & Stafford (2009), Dagher et al. (2016), DeAngelo, DeAngelo & Gilson (1994), Dell’Ariccia, Laeven & Suarez (2016), Domanski, Shin & Sushko (2015), Drechsler et al. (2016), Eber (2016), EIOPA (2014), Ellul et al. (2015), Esty (1998), Fama & French (1997), FIO (2013), Foley-Fisher et al. (2015, 2016), Gertler & Karadi (2011), Gertler & Kiyotaki (2010), Gilje (2016), Greenstone, Mas & Nguyen (2014), Hanson, Kashyap & Stein (2011), Hanson et al. (2015), Hartley, Paulson & Rosen (2016), He & Krishnamurthy (2013), Hellmann, Murdock & Stiglitz (2000), Holsboer (2000), IMF (2016), Jensen & Meckling (1976), Keeley (1990), Khwaja & Mian (2008), Kirti (2017), Koijen & Yogo (2015a, 2015b, 2016), Maggiori (2016), Marcus (1984), Matutes & Vives (2000), McDonald & Paulson (2015), Merrill et al. (2012, 2014), Mian & Sufi (2010a, 2010b), NAIC (2012, 2013), Plantin & Rochet (2007), Plosser & Santos (2014), Poterba (1997), Rajan (2006), Rampini, Viswanathan & Vuillemey (2015), Repullo (2004), Rochet (2008), Sarin & Summers (2016), Seifert (2014), Stiglitz & Weiss (1981), Thompson (2011), Zentefis (2014).

### A Data — Overview
- Data sources:
  - Insurance regulatory data from SNL Financial (sourced from NAIC regulatory filings).
  - Bond and issuer characteristics from Mergent FISD.
  - Interest rates and swap rates from Bloomberg.
- Sample scope:
  - Restrict attention to life insurance subsidiaries.
  - Focus on bonds held in the general account.
  - Data period: 2005-2014.
  - Data frequency: quarterly holdings and daily transactions.
- Aggregation and identifiers:
  - Aggregate to group level pro-forma based on current ownership relationships reported by SNL.
  - Companies identified by NAIC Company Codes; groups by NAIC Group Numbers.
  - Insurers not part of any insurance group retained and identified by NAIC Company Codes.

### A.1 Insurers — Data and construction
- Source schedules and reporting:
  - Bonds in the general account reported on Schedule D.
  - Derivatives reported on Schedule DB.
  - Aggregate data based on insurance regulatory filings (not SEC filings or annual reports).
- Coverage and filters:
  - Restrict to bonds held in the general account and life insurance subsidiaries.
  - Drop observations where fair value is 0, missing or negative, or where par value is 0.
  - Calculate price as ratio of fair value scaled by par value; drop observations where price is outside the interval [0.5,2].
  - Drop hybrid securities, or bonds with maturity date before the relevant quarter.
  - If no rate type reported, assume fixed rate.
  - Drop observations with missing coupon information, or if the bond has 0 maturity and a fixed rate.
  - Only calculate DV01 for fixed-rate bonds; set DV01 to 0 when bond explicitly categorized as floating rate, or reported exactly at par value.
- Bond characteristic matching and supplements:
  - SNL provides company investment key to match bonds to owners.
  - Bond characteristics include CUSIP, asset type, issuer type, maturity, coupon, rate type (fixed or floating).
  - Collect supplemental bond characteristics and offering date from Mergent FISD.
  - Use issuer’s SIC code from Mergent FISD to identify financial issuers based on Fama & French (1997) industry.
- Maturity rounding and exclusions:
  - Round bond maturity in a given quarter (as of reporting date at end of quarter) to the nearest half year.
  - Drop observations where reported fair value not within 1% of estimated fair value (after yield-to-maturity matching).

### A.1.1 General account bonds — Pricing and risk measures
- Price and yield-to-maturity calculations:
  - Fair value reported at the end of each year when held; for interim quarters fair value is a blended value based on end-of-previous-year fair value, price paid for bonds bought and consideration received for bonds sold since the end of the previous year.
  - Calculate yield to maturity (Y) numerically using bond price formula:
    - P[N, C, f, Y] = sum_{i=1}^{f N} (C/f) / (1 + Y/f)^i + 1 / (1 + Y/f)^{f N} = (1 + Y/f)^{-f N} + C/Y [1 − (1 + Y/f)^{-f N}] = C/Y + (Y − C) / (1 + Y/f)^{f N}
  - Assuming semi-annual payment (f = 2), use a bisection search to calculate the yield to maturity.
  - Run 20 iterations of a bisection search and use the final midpoint as the yield to maturity.
  - Match estimated price using calculated yield to maturity to reported fair value; drop if reported fair value not within 1% of estimated fair value.
- DV01 and duration:
  - Calculate DV01 at the bond identifier level based on prices reported by each insurer by quarter.
  - Calculate DV01 numerically as the difference in price given a 1% shift in yield to maturity.
  - Calculate duration assuming annual coupon payments.
- Asset class classification rules:
  - Corporate bonds: issuer reported by SNL as industrial or utility, issuer not in Fama & French (1997) finance industry, and bond not structured.
  - Government bonds: issued by US Federal government, a government agency, US states, local governments or foreign governments, excluding any structured bonds.
  - Financial firms: bonds issued by financial firms (Fama & French (1997) finance industry) excluding structured bonds.
  - Private structured bonds: bonds issued by an industrial or utility issuer and are structured (includes RMBS, CMBS, MBS, ABS and other structured bonds).
  - Agency bonds: structured bonds issued by any type of government issuer.

### A.1.2 Derivatives outstanding — Interest-rate swaps
- Coverage and reporting changes:
  - Data on interest-rate derivatives for 2005-2014 on a quarterly basis.
  - Filter for standard single-currency interest-rate swaps.
  - Reporting requirements for derivatives changed in 2010, but swaps were consistently reported.
  - Prior to 2010: data reported on Schedule DB – Part C Section 1 (all collars, swaps and forwards open as of reporting date).
  - After 2010: data reported on Schedule DB – Part A Section 1 (all options, caps, floors, collars, swaps and forwards open as of reporting date).
  - Variables include NAIC Company code, text description, terms or strike (text field), notional amount and fair value; data is at transaction level; initiation and maturity dates reported for transactions.
- Initial filtering and exclusions:
  - Filter derivatives to retain observations likely to pertain to standard single-currency interest-rate swaps.
  - Use row numbers to perform initial screen and drop observations when row numbers not in ranges corresponding to designated interest-rate swap type.
  - Row number ranges retained:
    - Prior to 2010: hedging (500000-599999) and other (700000-799999).
    - After 2010: hedging effective (850000-859999), hedging other (910000-919999), replication (970000-979999), income generation (1030000-1039999) and other (1090000-1099999).
  - After 2010, keep observations for which risk type matches interest-rate risk.
  - Exclude transactions likely referring to CDS, currency swaps, total return swaps, inflation swaps or other miscellaneous swaps using description and strike fields.
  - Drop observations where initiation or maturity not reported.
  - For some insurers where two legs of swaps reported as separate line items, match legs based on row numbers and terms (counterparty, notional, initiation and maturity) and combine them.
- Parsing swap direction and fixed rate:
  - Parse direction and fixed rate on the fixed leg from text fields (description and strike).
  - For some firms, terms clearer in the description than in the strike.
  - Adjustments on floating leg (e.g., LIBOR + 50bps) are subtracted from fixed leg and treat swap as otherwise standard.
  - Merge in LIBOR and maturity-appropriate swap rate from Bloomberg for both initiation and reporting date.
  - Log-linearly interpolate the swap curve using the largest subset of maturities available from the set {2,3,5,7,10,20,30}.
  - Use interpolated curve to determine likely sign of fair value as of reporting date for a fixed-rate swap initiated on initiation date.
- Hierarchy for identifying direction and fixed rate:
  1. Best case: direction clearly identified and two separate interest rates identified from strike. Before categorizing, check whether sign of swap fair value matches expectation given market rates, unless fixed rate of swap was more than 1% different from swap rate at initiation, or the direction is very clearly reported.
  2. Next case: direction clearly identified but only a single rate can be identified from terms. Before using single rate as fixed rate, check whether it is closer to LIBOR as of reporting date than swap rate at initiation. Before categorizing, check whether sign of swap fair value matches expectation given market rates, unless fixed rate was more than 1% different from swap rate at initiation, or direction is very clearly reported.
  3. When information on direction is available but no information on the rate, use the swap curve at initiation to determine the fixed rate.
  4. When two rates are not available and the swap is not matched to this point, identify direction from sign of swap’s fair value in combination with changes in market conditions, and identify the fixed rate from the swap curve at initiation.

*Source: wp17180 - References*

### 5.  When two rates are available (in formats designating which rate is paid and which is received)

### 5.  When two rates are available (in formats designating which rate is paid and which is received)

### Swap identification and classification
- Procedure for ambiguous swaps:
  - If the swap is not matched to a point, check which rate is closer to LIBOR at the reporting date and which leg is closer to the swap curve at initiation.
- Summary of swap sample by identification type (as presented):
  - Type# of swaps (000’s)
  - 1399
  - 211
  - 317
  - 421
  - 50.2
  - Total448
- Observed pattern:
  - The majority of swaps are classified based on the first step.
  - Remaining steps are important; reporting is less clear for certain firms (e.g. Prudential and Aegon), making it more likely their swaps are classified in later steps.

### Swap valuation and DV01 calculation
- After identifying the direction and fixed rate on the fixed leg:
  - The fair value of a pay-fixed (receive-floating) swap is:
    - N otional×(1−P[N, C, f, Y])(2)
    - Note: the floating leg is always valued at par; the price of the bond is defined by Equation 1 in the source.
  - The price of a pay-floating swap is this value multiplied by -1.
- Yield curve and coupon assumptions:
  - Use the log-linearly interpolated swap curve as of the reporting date as the yield to price swaps.
  - Assume semi-annual payment of coupons.
- DV01 calculation:
  - DV01 is calculated numerically as the change in fair value given a one basis point shift in the swap curve.

### Bond transactions (general account) — data sources and construction
- Data availability and sources:
  - Bond acquisitions: Schedule D – Part 3 (transaction dates, purchase price, par value, NAIC Company Code, bond CUSIP).
  - Bond disposals: Schedule D – Part 4 (consideration received, par value).
  - Transaction-level data collected for 2005-2014.
  - Match transaction data to bond characteristics reported by SNL at NAIC Company Code and CUSIP level.
- Yield to maturity and DV01 for transactions:
  - Calculated numerically using the reported transaction price (or consideration received) scaled by par value.
  - Transactions reported exactly at par are treated separately (assume only par value is reported correctly); for these, match to DV01 per unit of par value from the quarterly view for the previous quarter (matching NAIC Company Code and CUSIP). If no match in previous quarter, match to current quarter.
- Weighted yield to maturity of purchases (insurer-quarter, by NAIC Group):
  - Calculate weighted average (weighting by fair value) of all purchases in the quarter.
  - Exclude bonds reported exactly at par from this calculation.
  - Use only prices reported on the day of the transaction, in the relevant quarter.
- Focus on newly issued bonds:
  - Identify newly issued bonds using issuance data from Dealogic (tranche level on all USD, fixed rate, corporate and private MBS bonds issued from 2005-2014, excluding money market securities).
  - Match issuance data to daily transaction data via CUSIPs.
  - Find first transaction date for a given CUSIP across all life insurers.
  - Restrict regressions to bonds where first transaction is within two weeks of the Dealogic deal pricing date.
  - Calculate fair value value weighted yield to maturity for all transactions by life insurers within three months of the first transaction date; regressions winsorize this variable at the 5th and 95th percentile by NAIC category.
  - Calculate fraction of net purchases within three months of first transaction date by insurers hit hard, as fraction of net purchases within three months by all large insurers with matched parents; regressions winsorize at 5th and 95th percentile by NAIC category.
  - Restrict regressions to bonds with at least ten transactions within this window by large insurers with matched parents.
  - Regressions include month dummies based on first transaction date, and issuer dummies based on Dealogic’s issuer classification (Issuer short code).

### Identification of passive sales and paydowns
- MBS paydowns:
  - MBS paydowns happen regularly on the 1st of December of each year and (to a lesser degree) the 1st of the last month of other quarters.
  - Classify any sales of structured bonds on these dates as paydowns.
- Regular cycles for other bonds:
  - Find mode sale month and mode sale day for a CUSIP.
  - If sales within one day (mode day ±1, e.g. 14th-16th) in an annual, semi-annual or quarterly cycle account for half of the number of sales, and more than 50 sales are reported for the CUSIP, identify sales fitting the pattern as passive sales.

### Derivatives transactions and DV01 changes
- Data elements:
  - All derivatives positions reported open at the end of each reporting period provide initiation dates.
  - All transactions terminated within a year are reported separately, allowing identification of terminations and transactions held only within a quarter.
- Reporting schedules:
  - Prior to 2010: Schedule DB – Part C Section 3 (all collars, swaps and forwards terminated during current year).
  - After 2010: Schedule DB – Part A Section 2 (all options, caps, floors, collars, swaps and forwards terminated during current year).
- Use quarterly data for initiations when possible.
- Proceed to find changes in DV01 resulting from transactions as above.

### Banks — data and variable construction
- Bank holding company data:
  - Form FR-Y9C for 2005Q1-2014Q4 from the Chicago Federal Reserve.
  - Aggregate data pro-forma to present ownership using data on bank holding company mergers (Chicago Federal Reserve).
  - Match banks to present ownership and check for chains of mergers or acquisitions as in Kirti (2017).
  - Restrict sample to banks with at least $10BN in assets as of 2014Q4.
- Call report data:
  - Commercial bank level call reports for 2001Q1-2014Q4 from the FFIEC website.
  - Aggregate call report data to parent holding company on a pro-forma basis using current ownership reported by SNL.
- Loan growth measures (holding company level, changes in natural logarithms):
  - Real estate loans: bhck1410 (loans secured by real estate), reported on Schedule HC-C. Allowances for loan and lease losses are not deducted on Schedule HC-C.
  - Total loans: sum of bhckb529 (loans and leases, net of unearned income and allowance for loan and lease losses) and bhck5369 (loans and leases held for sale), reported on Schedule HC. This definition deducts the allowance for loan and lease losses.
  - Non real estate loans: total loans minus real estate loans, where total loans are inflated by the total allowance for loan and lease losses (bhck3123).
  - Therefore:
    - Real estate loan growth and non-real estate loan growth are before loan and lease losses are deducted.
    - Total loan growth is after loan and lease losses are deducted.
- Fraction of loans with maturity greater than one year:
  - Based on call report data on Schedule RC-C Part I: rcona564-rcona569 and rcfda570-rcfda575.
  - Refers to fraction of loans with reported breakdowns where remaining maturity or next repricing date is more than one year.

### Figures and table highlights (notes preserved)
- Figure 1: Composition of assets for US life insurers and banks (2014)
  - Uses aggregate Flow of Funds data as of 2014.
  - All values are in trillions of dollars.
- Figure 2: Net DV01 gap for large insurers with matched parents
  - Quarterly data from 2005Q1-2014Q4.
  - Bond DV01, Derivatives DV01 and the gap to Liability DV01 shown stacked (all in $BN).
  - Bond and derivatives DV01 shown scaled as positive even though they are negative numbers.
  - Liability DV01 estimated assuming a constant modified duration of 15 years.
- Figure 3: Insurers’ interest-rate swaps: total, net receive fixed, and duration-equivalent notional
  - Shows measures in $BN for 2010Q2, 2012Q2 and 2014Q4.
  - Duration-equivalent notional: amount of bonds insurers would need to hold (with same duration as bond portfolio) to obtain equivalent duration exposure.
- Figure 4: Risk taking based on crisis experience for large insurers with matched parent
  - Quarterly data from 2005Q1-2014Q4.
  - Panel A: Interest-rate risk (Net DV01 Gap, in percentage points).
  - Panel B: Credit risk (average YTM on bonds purchased, in percentage points).
  - Cutoffs: pre-crisis to crisis period at 2007Q2; crisis to post-crisis period at 2010Q4.
- Figure 5: Net MBS purchases as a fraction of total net purchases from 2005-2014
  - Uses daily transaction level data from 2005-2014.
  - Purchases shown as cumulative net purchases (based on reported fair values), as fraction of total net purchases over full sample period, value weighted and in percentage points.
  - Prepayments excluded. Privately issued structured bonds and Agency MBS shown separately.
- Figure 6: Private structured bonds bought by insurers hit hard in 2005 by rating
  - Rating distribution based on NAIC categories reported in insurance regulatory data.
  - Ratings shown at end of 2005 and end of 2009.
  - Weights (in percentage points) based on net purchase shares in 2005.
  - NAIC 1 omitted (can be inferred from sum of bars).
- Figure 7: Risk taking based on crisis experience for large banks with matched parent
  - Quarterly data from 2005Q1-2014Q4.
  - Panel A: Interest-rate risk (fraction of loans maturing or repricing in more than one year, in percentage points).
  - Panel B: Credit risk (log changes in real estate loans, multiplied by 100).
  - Cutoffs: 2007Q2 and 2010Q4.

### Selected table summary statistics (as presented)
- Table 1: Summary statistics for large insurers with matched parent (quarterly, 2005Q1-2014Q4)
  - Variables and values shown (means or counts) exactly as in source:
    - Assets (BN) 92.3 38.6 135.8 113.7
    - Equity/Assets 8.0 9.6 6.8 4.8
    - Bond DV01 Gap 42.1 39.2 44.5 20.5
    - Net DV01 Gap 40.3 39.0 41.4 20.9
    - YTM (Purchases) 470.9 484.7 459.7 112.7
    - Dur (Purchases) 10.0 9.6 10.3 2.1
    - Insurer-Quarters 1,701
    - Insurers 50
- Table 2: Regressions of Net DV01 Gap and YTM on crisis experience
  - Net DV01 Gap definition: (Liability DV01 - Bond DV01 - Derivatives DV01)/(Liability DV01), in percentage points with a cap at 100 and floor at 0.
  - YTM definition: fair value value weighted average yield to maturity of all bonds purchased in insurer-quarter, in basis points, winsorized at the 5th and 95th percentiles.
  - Crisis Hit Flag: insurer-level dummy for severe dividend cuts, reduction in equity/assets ratio or equity issuance during the crisis period (as described in Section 3).

*Source: wp17180 - 5.  When two rates are available (in formats designating which rate is paid and which is received).*

### Section  3.   All  specifications  include  interactions  of  Crisis  Hit  Flag  with  dummies  for  the  crisis

### wp17180 - Section  3.   All  specifications  include  interactions  of  Crisis  Hit  Flag  with  dummies  for  the  crisis

### Share of newly issued corporate bonds bought by insurers hit hard (Table 3)
- Sample and methodology:
  - Daily transaction data from 2005-2014, corporate bonds (excluding financial issuers).
  - Regressions at the issue level. Dependent variable: net purchases by insurers hit hard within three months as a fraction of net purchases by all large insurers with matched parents, in percentage points, winsorized at the 5th and 95th percentiles by NAIC category.
  - Sample restricts to bonds with at least ten transactions by large insurers with matched parents.
  - YTM: fair value weighted yield to maturity for life insurers within three months, winsorized at the 5th and 95th percentiles by NAIC category.
  - Maturity from Dealogic. Controls include (origination) month and issuer fixed effects, (log) total purchases, dummy for private structured bonds, and interactions with crisis and post-crisis dummies.
  - Standard errors double-clustered by issuer and month.
- Key coefficients (All IG):
  - YTM: 12.0 (t-statistic 4.17)
  - YTM×2007Q3-2010Q4: -14.9 (t-statistic -5.31)
  - YTM×2011Q1-2014Q4: -17.7 (t-statistic -5.91)
  - Maturity: -0.4 (t-statistic -2.75)
  - Maturity×2007Q3-2010Q4: 0.8 (t-statistic 4.95)
  - Maturity×2011Q1-2014Q4: 0.9 (t-statistic 5.71)
- Fit and sample:
  - R 2: 0.57
  - Issues: 3,058
  - Issuers: 837
  - Transactions: 66,056

### Share of newly issued private investment grade bonds bought by insurers hit hard (Table 4)
- Sample and methodology:
  - Daily transaction data from 2005-2014, corporate bonds including financial issuers and private structured bonds.
  - Dependent variable and controls analogous to Table 3; standard errors double-clustered by issuer and month.
- Key coefficients (All IG):
  - YTM: 3.5 (t-statistic 1.04)
  - YTM×2007Q3-2010Q4: -5.6 (t-statistic -1.66)
  - YTM×2011Q1-2014Q4: -9.5 (t-statistic -2.80)
  - Maturity: -0.3 (t-statistic -2.00)
  - Maturity×2007Q3-2010Q4: 0.7 (t-statistic 4.42)
  - Maturity×2011Q1-2014Q4: 0.8 (t-statistic 5.16)
- Fit and sample:
  - R 2: 0.55
  - Issues: 4,144
  - Issuers: 1,205
  - Transactions: 86,778

### Interest-rate risk (with and without derivatives) based on crisis experience for large insurers (Table 5)
- Sample and methodology:
  - Quarterly data from 2005Q1-2014Q4, aggregated to insurance group level pro-forma.
  - DV01 Gap including derivatives = (Liability DV01 - Bond DV01 - Derivatives DV01)/(Liability DV01), percentage points, cap at 100 and floor at 0.
  - DV01 Gap excluding derivatives = (Liability DV01 - Bond DV01)/(Liability DV01), percentage points, cap at 100 and floor at 0.
  - Crisis Hit Flag: insurer-level dummy for severe dividend cuts, reduction in equity/assets ratio or equity issuance during crisis (as described in Section 3).
  - All specifications include interactions of Crisis Hit Flag with crisis (2007Q3-2010Q4) and post-crisis (2011Q1-2014Q4) dummies; Log(Assets) included. Standard errors double-clustered by insurer and quarter.
- Key coefficients (Including derivatives):
  - Crisis Hit Flag: 6.85 (t-statistic 1.06)
  - Crisis Hit Flag×2007Q3-2010Q4: -3.22 (t-statistic -2.63)
  - Crisis Hit Flag×2011Q1-2014Q4: -8.96 (t-statistic -2.74)
  - Log(Assets): 0.29 (t-statistic 0.12)
  - R 2: 0.09
  - Insurer-Quarters: 1,701
  - Insurers: 50
- Key coefficients (Excluding derivatives):
  - Crisis Hit Flag: 6.22 (t-statistic 0.96)
  - Crisis Hit Flag×2007Q3-2010Q4: -2.94 (t-statistic -1.26)
  - Crisis Hit Flag×2011Q1-2014Q4: -8.07 (t-statistic -2.20)
  - Log(Assets): -0.31 (t-statistic -0.05)
  - R 2: 0.83
  - Insurer-Quarters: 1,701
  - Insurers: 50

### Yields and duration of purchases based on crisis experience for large insurers with matched parent (Table 6)
- Sample and methodology:
  - Quarterly data from 2005Q1-2014Q4, aggregated to insurer-group level pro-forma.
  - YTM: fair value weighted average yield to maturity of all bonds purchased in insurer-quarter, in basis points, winsorized at the 5th and 95th percentiles.
  - Duration: fair value weighted average duration of all bonds purchased in insurer-quarter, in years, winsorized at the 5th and 95th percentiles.
  - Crisis Hit Flag and interactions as in Table 5. Log(Assets) included. Standard errors double-clustered by insurer and quarter.
- Key coefficients (YTM / Credit risk):
  - Crisis Hit Flag: 13.64 (t-statistic 1.15)
  - Crisis Hit Flag×2007Q3-2010Q4: -28.28 (t-statistic -2.45)
  - Crisis Hit Flag×2011Q1-2014Q4: -16.19 (t-statistic -1.08)
  - Log(Assets): -19.26 (t-statistic -2.85)
  - R 2 (YTM): 0.63
- Key coefficients (Duration of purchases / Interest-rate risk):
  - Crisis Hit Flag: 0.52 (t-statistic 1.21)
  - Crisis Hit Flag×2007Q3-2010Q4: 0.69 (t-statistic 1.90)
  - Crisis Hit Flag×2011Q1-2014Q4: -0.15 (t-statistic -0.30)
  - Log(Assets): 0.02 (t-statistic 0.15)
  - R 2 (Duration): 0.12
- Sample:
  - Insurer-Quarters: 1,701
  - Insurers: 50

### Risk taking based on crisis experience for all large insurers (Table 7)
- Sample and methodology:
  - Quarterly data from 2005Q1-2014Q4, aggregated to insurance group level. Includes insurers without matched parents (classified as not hit hard).
  - Net DV01 Gap = (Liability DV01 - Bond DV01 - Derivatives DV01)/(Liability DV01), percentage points, cap at 100 and floor at 0.
  - YTM as in Table 6. Crisis Hit Flag and interactions as in prior tables. Log(Assets) included. Standard errors double-clustered by insurer and quarter.
- Key coefficients (Interest-rate risk, Net DV01 Gap):
  - Crisis Hit Flag: 1.47 (t-statistic 0.30)
  - Crisis Hit Flag×2007Q3-2010Q4: -1.89 (t-statistic -1.82)
  - Crisis Hit Flag×2011Q1-2014Q4: -7.73 (t-statistic -2.64)
  - Log(Assets): 2.61 (t-statistic 1.29)
  - R 2 (Net DV01 Gap): 0.10
- Key coefficients (Credit risk, YTM):
  - Crisis Hit Flag: 9.92 (t-statistic 1.01)
  - Crisis Hit Flag×2007Q3-2010Q4: -20.77 (t-statistic -1.81)
  - Crisis Hit Flag×2011Q1-2014Q4: -17.22 (t-statistic -1.26)
  - Log(Assets): -13.21 (t-statistic -2.17)
  - R 2 (YTM): 0.60
- Sample:
  - Insurer-Quarters: 2,136
  - Insurers: 64

### Summary statistics for large banks with matched parent (Table 8)
- Sample and variables:
  - Quarterly data from 2005Q1-2014Q4, aggregated to parent level pro-forma. Assets in $BN at holding company level. Loan growth = changes in natural logarithms × 100. Loans>1 Year = loans maturing or repricing in more than one year as a fraction of loans with reported maturity data, in percentage points, using call report data at commercial bank level.
  - ‘Hit hard’ bank-level dummy for severe dividend cuts or reduction in equity/assets ratio (as described in Section 3).
- Key descriptive statistics (Full Sample / Others / Hit hard / Std. Dev.):
  - Assets: 184.0 / 148.3 / 226.0 / 473.6
  - Loan Growth: 1.1 / 1.2 / 1.0 / 2.7
  - RE Loan Growth: 0.7 / 1.0 / 0.4 / 2.8
  - Non-RE Loan Growth: 1.4 / 1.5 / 1.2 / 6.2
  - Loans>1 Year: 47.1 / 48.2 / 45.7 / 17.2
  - BHC-Quarters: 2,144
  - BHCs: 54

### Risk taking for large banks with matched parent — RE loan growth and loan maturity (Table 9)
- Sample and methodology:
  - Quarterly data from 2005Q1-2014Q4, aggregated to parent level. RE loan growth = changes in natural logarithms of real estate loan growth × 100. Loans>1 Year as in Table 8. Crisis Hit Flag and interactions as above. Log(Assets) included. Standard errors double-clustered by bank and quarter.
- Key coefficients (Credit risk, ∆log×100 RE loan growth):
  - Crisis Hit Flag: 0.31 (t-statistic 0.87)
  - Crisis Hit Flag×2007Q3-2010Q4: -1.02 (t-statistic -2.47)
  - Crisis Hit Flag×2011Q1-2014Q4: -0.75 (t-statistic -1.56)
  - Log(Assets): -0.31 (t-statistic -3.45)
  - R 2 (RE loan growth): 0.22
- Key coefficients (Interest-rate risk, Loans>1Yr):
  - Crisis Hit Flag: 2.23 (t-statistic 0.57)
  - Crisis Hit Flag×2007Q3-2010Q4: -1.04 (t-statistic -2.41)
  - Crisis Hit Flag×2011Q1-2014Q4: -0.61 (t-statistic -1.24)
  - Log(Assets): 1.17 (t-statistic 2.11)
  - R 2 (Loans>1Yr): 0.37
- Sample:
  - BHC-Quarters: 2,144
  - BHCs: 54

### Additional measures of loan growth for large banks (Table 10)
- Sample and methodology:
  - Quarterly data from 2005Q1-2014Q4, aggregated to parent level. Credit growth = changes in natural logarithms of loan growth × 100. Total loans from Schedule HC net of allowance; components from Schedule HC-C without deducting allowance. Crisis Hit Flag and interactions as above. Log(Assets) included. Standard errors double-clustered by bank and quarter.
- Key coefficients (Non-real estate loans, ∆log×100):
  - Crisis Hit Flag: 0.44 (t-statistic 0.95)
  - Crisis Hit Flag×2007Q3-2010Q4: -1.35 (t-statistic -2.26)
  - Crisis Hit Flag×2011Q1-2014Q4: -0.40 (t-statistic -0.81)
  - Log(Assets): -0.16 (t-statistic -1.14)
  - R 2: 0.05
- Key coefficients (All loans, ∆log×100):
  - Crisis Hit Flag: 0.22 (t-statistic 0.71)
  - Crisis Hit Flag×2007Q3-2010Q4: -1.41 (t-statistic -2.22)
  - Crisis Hit Flag×2011Q1-2014Q4: -0.10 (t-statistic -0.13)
  - Log(Assets): 3.42 (t-statistic 1.69)
  - R 2: 0.11
- Key coefficients (Additional All loans specifications):
  - Crisis Hit Flag: 0.44 (t-statistic 0.95) and 0.22 (t-statistic 0.71) across columns
  - Crisis Hit Flag×2007Q3-2010Q4: -0.74 (t-statistic -2.29) and -0.77 (t-statistic -2.21)
  - Crisis Hit Flag×2011Q1-2014Q4: -0.22 (t-statistic -0.59) and -0.17 (t-statistic -0.42)
  - Log(Assets): -0.19 (t-statistic -1.94) and 0.76 (t-statistic 2.02)
  - R 2: 0.21 and 0.35
- Sample:
  - BHC-Quarters: 2,144
  - BHCs: 54

*Source: wp17180 - Section  3.   All  specifications  include  interactions  of  Crisis  Hit  Flag  with  dummies  for  the  crisis*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2017/wp17180.pdf_