## 1. Ergodic distributions

## Source details

**Canonical URL:** [1. Ergodic distributions](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2010/_wp10164.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2010/_wp10164.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2010/_wp10164.pdf.json)

---

### Ergodic distributions — key findings
- Figure 7 displays “conjectured” ergodic distributions of b for dates 1, 8, 40, 41 and 48 in the BL model and the true ergodic distribution of the RE model.
- The actual ergodic distribution of the BL model equals that of the RE model because agents learn F^a in the long run and the long-run equilibrium coincides with rational expectations.
- Conjectured ergodic distributions are agents’ projections of the long-run equilibrium conditional on their current beliefs about the dynamics of the financial regime; plotting conjectured and RE long-run distributions illustrates how optimism and pessimism affect willingness to borrow or save.
- Appendix B provides computation details for these long-run distributions.

### Period-specific comparative observations (agents’ conjectured long-run distributions)
- Period 1:
  - Mean of bond holdings pre-financial-innovation: −0.31.
  - Agents conjecture the support of the long-run distribution of bonds will shift sharply to the left (supporting higher debt levels) and project much lower saving than eventual new stochastic steady state.
  - RE ergodic distribution shows typical bimodal shape of a two-point regime-switching process with high persistence.
- Period 8:
  - Agents conjecture debt positions in the [−0.54, −0.49] range as most likely long-run equilibria, while the RE ergodic distribution assigns zero probability to that range.
- Period 40 (peak of optimism):
  - Conjectured long-run mass concentrated on large debt ratios in [-0.59, −0.54], which are never long-run equilibria under RE.
  - Overborrowing and undersaving relative to RE optimality peak at date 40.
- Period 41 onward and period 48 (pessimistic phase):
  - After first κ_l realization, agents reverse: they “underborrow” and “oversave,” with peak pessimism by date 48.
  - By period 48 agents’ conjectured distribution is two-peaked because they have learned shifts to κ_h are possible.

### Interpretive summary
- Short sample realizations of κ_h suffice for large optimism relative to RE transition probabilities, shifting conjectured long-run distributions toward higher debt; subsequent κ_l realizations reverse beliefs and produce opposite shifts.

---

### 2. Forecast functions

### Construction and methodology
- Forecast functions trace expected values of endogenous variables along the AU recursive equilibrium path for dates 0 to 48 using:
  - The sequence of beliefs {f(F_s | κ_t)} and corresponding date-t decision rules (AUOP solutions).
  - A law of motion for the probability of the economy being in each triple (b, z, κ) conditional on the particular κ_T history (not averaged across κ realizations).
- Initial conditions focused on the triple approximating the U.S. economy in 1996Q4:
  - Date 1 initial conditions: (b1, z1, κ_h) with b1 = −0.345 and z1 = 1.
  - κ sequence: κ_h for t = 2, ..., 40 and κ_l for t = 41, ..., 48.
- Policy function b′_t(b, z, κ) is time-varying because it depends on date-specific beliefs.

### Forecast outputs and model comparisons (Figure 8)
- Forecasts plotted as percent deviations from long-run means for:
  - bond holdings as a share of output (b′/y),
  - consumption,
  - price of land,
  - savings rate (GSF/y).
- Long-run means identical across learning and RE models (ergodic distributions coincide).
- Line styles: solid (blue) BL model; dashed (green) FVL model; dotted (red) RE model.

### Bonds — comparative dynamics (Panel (a))
- During the optimistic phase the BL model exhibits greater borrowing than RE and FVL.
- Initial two periods after financial innovation: three models predict similar debt dynamics.
- By period 40 (optimism peak) bond holdings in BL decline by as much as 21 percentage points below the long-run average.
- BL displays much larger decline in bond holdings due to optimism and debt-deflation feedback; RE and FVL remain similar to each other.

---

### 3. Turning points and aggregate comparisons

### Mechanisms: learning + debt-deflation vs alternatives
- BL (learning + price-dependent collateral) produces a much stronger decline in assets and larger corrections than:
  - RE: agents know true F^a (no learning).
  - FVL: Bayesian learning but collateral valued at fixed long-run q̄ l (removes credit externality and Fisherian deflation).
- Date-41 switch to κ_l behaves like an unexpected shock because by date 40 agents view κ_h as nearly absorbent.

### Quantitative turning-point comparisons (Table 2 reported values)
- Timing: period 0 = 1996Q4; period 40 = 2006Q4; period 48 = 2008Q4.
- Peak of optimism (t = 40) — differences relative to period 0:
  - Data: E[(b/y)40 − (b/y)0] = −0.355
  - RE: −0.083
  - FVL: −0.089
  - BL: −0.223
  - Data: E[(qL/y)40 − (qL/y)0] = 0.280
  - RE: −0.025
  - FVL: 0.305
  - BL: 0.122
- Financial crisis (t = 48) — differences relative to peak t = 40:
  - Data: E[(b/y)48 − (b/y)40] = 0.023
  - RE: 0.122
  - FVL: 0.133
  - BL: 0.254
  - Data: E[(qL/y)48 − (qL/y)40] = −0.149
  - RE: 0.013
  - FVL: −0.305
  - BL: −0.121

### Key turning-point findings
- BL accounts for 63 percent of the rise in net household debt during 1997–2006 (model b/y falls by almost 23 percentage points vs. 36 percentage points in data).
- BL accounts for 44 percent of the rise in residential land prices during 1997–2006 (model q increases almost 13 percentage points vs. 28 percentage points in data).
- BL generates a large debt reversal of about 25 percentage points between t = 40 and t = 48; data reversal was 2.3 percentage points (model overstates reversal partly because model bonds are one-period whereas real-world debt has longer maturity).
- BL matches observed land price decline during crisis reasonably: model decline 12.1 percentage points vs. data decline 14.9 percentage points.

---

### 10. Excess Returns — conceptual mechanism and calibration

### Pre-crisis facts and financial innovation context
- 1996–2006 household net credit market assets: fell from −35 to −70 percent of GDP.
- Market value of residential land / GDP: rose from 45 to nearly 75 percent between 1996 and 2006.
- Macro leverage measure (net credit market assets / market value of residential land): rose from 0.64 to 0.93 in absolute value during 1996–2006.
- Key dates and markets:
  - First CDOs: early 1980s; securitization boom mid 1990s with CMOs.
  - CDSs on CMO payments introduced by mid 2000s; by end-2007 CDS market ≈ $45 trillion (or 3 times U.S. GDP).
- Regulatory changes:
  - 1995 New Community Reinvestment Act (strengthened Fannie Mae and Freddie Mac).
  - 1999 Gramm–Leach–Bliley Act.
  - 2000 Commodity Futures Modernization Act.
- Financial conditions pre-crisis:
  - Fraction of banks tightening credit standards jumped from nearly zero to over 50 percent at crisis onset.
  - Median downpayment on conventional mortgages rose from 5 to 13 percent (Feb. 16, Wall Street Journal).

### Model environment and borrowing constraint
- Assets: a non-state-contingent bond and land in fixed supply.
- Borrowing constraint: b_{t+1}/R ≥ −κ_t q_t l_{t+1}, where κ_t is regime-dependent leverage fraction.
- Financial innovation modeled as introduction of high-leverage regime κ_h; economy can switch to low-leverage κ_l.

### Learning and AU solution strategy
- Agents know κ_h and κ_l but not true transition probabilities F^a; they are Bayesian learners observing realizations of κ without noise and updating posteriors.
- Anticipated Utility (AU) two-stage solution:
  - (i) compute posterior sequence {f(F^s | κ_t)} by Bayesian updating;
  - (ii) solve date-t AU optimization problems conditional on date-t posterior to obtain policy and pricing functions; chain these to get equilibrium dynamics.

### Asset pricing and exact excess return expression
- Land price q_t equals expected discounted sum of future land dividends with stochastic discount factors adjusted for Lagrange multiplier μ on the credit constraint (shadow value).
- Exact expression preserved for excess return:
  - E_s_t[R^q_{t+1} − R] = ((1−κ_t) μ_t − cov_s_t(λ_{t+1}, R^q_{t+1})) / E_s_t[λ_{t+1}],
  - where R^q_{t+1} ≡ (z_{t+1} g′(l_{t+1}) + q_{t+1})/q_t and λ_{t+1} ≡ β u′(c_{t+1}).
- Channels:
  - Direct effect: (1−κ_t) μ_t raises excess return when constraint binds.
  - Indirect effects: collateral constraint hampers consumption smoothing, reduces cov_s_t(λ_{t+1}, R^q_{t+1}), and lowers E_s_t[λ_{t+1}], affecting required returns and valuations.

### Interaction and quantitative amplification
- Learning-driven mispricing inflates/deflates collateral values; inflated collateral permits more borrowing, amplifying booms; deflation tightens collateral and magnifies contractions.
- Quantitatively, effects on debt and asset prices are nearly twice as large when both debt-deflation mechanism and learning mechanism interact than when either mechanism is removed.

### Baseline calibration and experiment timing (explicit values preserved)
- Experiment dating:
  - Financial innovation begins 1997Q1; crisis begins 2007Q1.
  - Optimistic phase: κ_t = κ_h for t = 1, ..., J with J = 40 quarters (1997Q1–2006Q4).
  - κ_l realization in 2007Q1; κ_t = κ_l for t = J+1, ..., T with T = 48 quarters (1997Q1–2008Q4).
- Baseline calibrated parameter values:
  - R = 2.66 percent.
  - β = 0.91.
  - σ = 2.0.
  - α = 0.0251.
  - L = 1.0.
  - c = 0.670.
  - A = 0.321.
  - ̄b = −0.313.
  - κ_l = 0.642.
  - κ_h = 0.926.
  - TFP Markov approximation: ρ = 0.869, σ_e = 0.00833 (Tauchen–Hussey with 9 states).
- Priors and calibration to excess returns:
  - Target excess return: Fannie Mae 30-year residential MBS spread in 1997Q2 = 47.6 basis points.
  - Calibration yields n_0 = 0.0205 and initial prior Beta(0.0205, 0.0205) (U-shaped, mean 0.5).
- True transition probabilities:
  - F^a_hh = 0.964, F^a_ll = 0.964 (symmetric true process with mean duration of κ_h ≈ 7 years).

### Baseline learning implications
- With Beta(0.0205,0.0205) priors, an initial long spell of κ_h drives E_t[F^s_hh] close to 1 within a few quarters, implying perceived mean durations of order 50–1,000 quarters and much-reduced coefficient of variation of κ.
- Agents become optimistic, reduce precautionary savings, increase leverage and land prices above RE benchmark; first κ_l realization triggers sharp downward belief update and abrupt contractions.

### Quantitative aggregate contributions
- Model accounts for 63 percent of rise in net household debt during 1997–2006.
- Model accounts for 44 percent of rise in residential land prices during 1997–2006.
- Comparative experiments:
  - BL: Bayesian learning + price-dependent collateral + debt-deflation.
  - RE: agents know F^a.
  - FVL: Bayesian learning but collateral valued at constant q̄ l (no credit externality / Fisherian deflation).
- Learning + credit externality produces pronounced boom and sharp corrections; interaction nearly doubles effects relative to one mechanism alone.

---

### 4. Projected Excess Returns on Land

### Projection setup
- Expected excess returns computed up to 50 periods ahead using beliefs and decision rules as of initial dates t = 1, 40, and 41.
- Initial state for projections:
  - b set at mean bond holdings from forecast function for corresponding date,
  - κ set to corresponding κ_t,
  - z set to mean value z = 1.
- Expectations conditional on date-specific equilibrium pricing function.

### Date-specific comparative findings
- Date 1 (panel a):
  - RE excess returns slightly exceed BL up to period 10; after period 10 ordering reverses and BL projects slightly higher returns.
  - FVL yields expected excess returns significantly below RE and BL.
  - Interpretation: FVL’s higher date-1 land price consistent with lower expected excess returns.
- Date 40 (panel b):
  - After 40 periods of κ_h, expected excess returns ahead of date 40 are significantly lower in BL and FVL than as of date 1.
  - BL and FVL produce sharply higher land prices at date 40 than date 1.
  - Projected returns: BL < RE; FVL < BL.
- Date 41 (panel c — immediate post-switch to κ_l):
  - Projected excess returns for period 42 very high in all models due to sharp tightening of borrowing constraint.
  - Direct effect via (1−κ_l) μ_41 strongest in BL, next RE, weakest FVL.
  - Interest rate premia ordering mirrors direct effect ordering.

### Dynamics after switch and long-run ordering
- Immediately post-tightening:
  - BL and RE: borrowing constraint projected not to bind for a couple of periods, then bind again as debt rebuilds.
  - FVL: constraint projected to remain binding but tightness reduced via debt adjustment; projected returns decline.
- Beyond initial adjustment (after ~10 periods):
  - Projected returns in BL exceed RE and FVL.
  - FVL projected returns sharply lower than others due to absence of debt-deflation channel.

### Specific numeric illustrative values (preserved)
- Projections computed for 50 periods ahead for t = 1, 40, 41.
- One-period-ahead expected excess return in period 41 (bottom panel) reported as:
  - 1418 percent in baseline,
  - 74 percent in RE,
  - 30 percent in FVL.
- Figure 10 legend q values:
  - Panel (a) Period 1: BL q = 0.045; FVL q = 0.480; RE q = 0.294.
  - Panel (b) Period 40: BL q = 0.598; FVL q = 0.781; RE q = 0.451.
  - Panel (c) Period 41: BL q = 0.478; FVL q = 0.662; RE q = 0.465.

---

### 5. Sensitivity Analysis

### Overview and general result
- Sensitivity focuses on turning point effects and examines:
  - Changes in initial priors (Table 3 / Scenarios (1)–(4)).
  - Changes in κ_h, κ_l, β, and R (Table 4 / Scenarios (1)–(4)).
- General finding: parameterizations that generate larger booms in optimistic phase also generate larger busts in crisis.

### Sensitivity to initial priors (Table 3 reported entries)
- Baseline: n0 = 0.0205 (Beta(0.0205,0.0205)).
- Scenario (1) uniform priors (n0 = 1):
  - E[(b/y)40 − (b/y)0]: −0.087
  - E[(qL/y)40 − (qL/y)0]: −0.020
  - E[(b/y)48 − (b/y)40]: 0.137
  - E[(qL/y)48 − (qL/y)40]: −0.006
  - Debt buildup smaller (9 percentage points v. 22 in BL); crisis effects much weaker.
- Scenario (2) lower initial counters (n0 = 0.01):
  - E[(b/y)40 − (b/y)0]: −0.243
  - E[(qL/y)40 − (qL/y)0]: 0.142
  - E[(b/y)48 − (b/y)40]: 0.274
  - E[(qL/y)48 − (qL/y)40]: −0.141
  - Larger booms and busts than BL.
- Scenario (3) nhh0 = nll0 = 0.54 (nhl0, nlh0 = 0.0205):
  - E[(b/y)40 − (b/y)0]: −0.222
  - E[(qL/y)40 − (qL/y)0]: 0.121
  - E[(b/y)48 − (b/y)40]: 0.253
  - E[(qL/y)48 − (qL/y)40]: −0.120
  - Results similar to BL.
- Scenario (4) informative pre-innovation κ_l (nll0 = 68, nhh0 = 1):
  - E[(b/y)40 − (b/y)0]: −0.221
  - E[(qL/y)40 − (qL/y)0]: 0.121
  - E[(b/y)48 − (b/y)40]: 0.254
  - E[(qL/y)48 − (qL/y)40]: −0.123
  - Turning-point dynamics similar to BL because forecast functions condition on same κ histories.

### Sensitivity to other parameters (Table 4 reported entries)
- Scenarios change one parameter each:
  - Scenario (1): R = 1.0098 (average interest rate 1997–2008) vs BL R = 2.66 percent.
  - Scenario (2): κ_l = 0.75.
  - Scenario (3): κ_h = 0.8.
  - Scenario (4): β = 0.95 (vs BL β = 0.91).
- Reported values:
  - E[(b/y)40 − (b/y)0]:
    - BL −0.223; R = 1.0098 −0.248; κl = 0.75 −0.240; κh = 0.8 −0.143; β = 0.95 −0.237
  - E[(qL/y)40 − (qL/y)0]:
    - BL 0.122; R = 1.0098 0.158; κl = 0.75 0.140; κh = 0.8 0.120; β = 0.95 0.140
  - E[(b/y)48 − (b/y)40]:
    - BL 0.254; R = 1.0098 0.208; κl = 0.75 0.157; κh = 0.8 0.175; β = 0.95 0.175
  - E[(qL/y)48 − (qL/y)40]:
    - BL −0.121; R = 1.0098 −0.038; κl = 0.75 −0.059; κh = 0.8 −0.120; β = 0.95 0.098

### Notable parameter effects
- Lower R (0.98 percent) increases long-run RE means (E[b/y] from −0.36 to −0.41 and land price from 0.45 to 0.53) and generates larger optimistic-phase increases in debt and asset prices; crisis corrections smaller.
- Higher κ_l (0.75) yields slightly larger booms and smaller reversals.
- Lower κ_h (0.8) reduces size of buildup and reversal; land price boom/crash marginally affected.
- Higher β (0.95) supports higher asset prices and slightly larger booms; crisis effects on asset prices weaker and recovery quicker.

### Sensitivity interpretation
- Initial priors calibrated to observed excess returns (n0 = 0.0205) are important to discipline magnitude of effects.
- Priors can be set to generate larger or smaller booms and busts; a data-based calibration is necessary for relevance.
- Across parameter changes, larger optimistic-phase booms are associated with larger crisis-phase busts.

---

### Methods and computation (Appendices A and B — brief)
- Appendix A: AUOP recursive solution algorithm (six steps) iterating posteriors, date-t pricing functions, policy functions, and convergence checks; notes on passive Bayesian learning and speed-ups.
- Appendix B: computation of ergodic distributions and forecast functions under RE and AU forecast functions under learning; law of motion for distributions; formulas for expected returns j periods ahead and present-value reconstruction of q_t via one-period-ahead expected returns.

---

### Policy-relevant implications and recommendations (from analysis and conclusion)
- Financial innovation that materially alters leverage possibilities can lead to underpricing of risk when historical data are insufficient; Bayesian learning about regime persistence can generate prolonged optimism followed by abrupt pessimistic corrections.
- The collateral constraint linking borrowing to market valuations creates a potent amplification channel: overvaluation of collateral during optimistic phases feeds overborrowing; subsequent revaluation tightens constraints and magnifies downturns.
- Policy recommendations:
  - Close supervision of financial intermediaries in early stages of financial innovation is critical.
  - Monitoring and possibly attenuating rapid expansion in securitization and leverage—particularly when demonstrated performance histories are short—can reduce optimism-driven overborrowing.
  - Countercyclical tools that limit direct link between market valuations and collateral-based borrowing (or strengthen buffers against sharp revaluations) can mitigate debt-deflation amplification.
  - Taxes or capital requirements can address the credit externality component of overborrowing; however, they cannot fully correct overborrowing driven by optimistic beliefs if policymakers are as uninformed as households—nonetheless they reduce the overall magnitude (learning-only overborrowing ≈ one half of combined effect).

*Source: IMF Working Paper excerpt (sections: 1. Ergodic distributions; 2. Forecast functions; 3. Turning Points; 4. Projected Excess Returns on Land; 5. Sensitivity Analysis; 10. Excess Returns; Appendices A and B).*

### 1. Ergodic distributions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  24

### 1. Ergodic distributions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  24

### Major sections covered in this content unit
- 1. Ergodic distributions (page 24)
- 2. Forecast functions (page 25)
- 3. Turning Points (page 32)
- 4. Projected Excess Returns on Land (page 33)
- 5. Sensitivity Analysis (page 35)
- IV. Conclusion (page 39)
- References (page 41)
- Appendices (page 44)
  - A Solution Method (page 44)
  - B Computation of Ergodic Distributions, Forecast Functions, Excess Returns (page 45)

### Enumerated tables referenced (by table number and title)
- Table 1.   Baseline Parameter Values. (page 20)
- Table 2.   Average Changes at the Turning Points. (page 32)
- Table 3.   Sensitivity on Priors. (page 36)
- Table 4.   Sensitivity on Other Parameters. (page 38)

### Figures referenced (by figure number and title)
- Figure 1.   Net Credit Market Assets and Value of Residential Land. (page 4)
- Figure 2.   Timeline of Events During the Run-up to the U.S. Credit Crisis. (page 5)
- Figure 3.   Banks’ Willingness To Lend. (page 5)
- Figure 4.   Evolution of Beliefs. (page 15)
- Figure 5.   Beta Distributions. (page 22)
- Figure 6.   Evolution of Beliefs under Alternative Priors. (page 23)
- Figure 7.   Ergodic Distributions of Bond Holdings. (page 24)
- Figure 8.   Forecast Functions. (page 27)
- Figure 9.   Asset Price Dynamics. (page 31)

### Key thematic focus (as indicated by section and figure/table titles)
- Analysis and computation of ergodic distributions and their implications (section 1; Figure 7; Appendix B).
- Derivation and presentation of forecast functions (section 2; Figure 8; Appendix B).
- Identification and quantification of turning points, including average changes at the turning points (section 3; Table 2).
- Projection of excess returns on land (section 4).
- Sensitivity analysis with respect to priors and other parameters (section 5; Table 3; Table 4).
- Methodological details for solution and computation methods (Appendix A; Appendix B).
- Empirical/contextual illustrations relating credit, lending, beliefs, and asset-price dynamics (Figures 1–6, 9).

*Source: _wp10164 - 1. Ergodic distributions. . . . . . . . . . . . . . . . . . . . . . . . .  24*

### 10.  Excess Returns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  34

### 10.  Excess Returns

### Financial innovation, market structure, and pre-crisis facts
- Between 1996 and 2006 the net credit market assets of U.S. households fell from −35 to −70 percent of GDP.
- Between 1996 and 2006 the market value of residential land as a share of GDP rose from 45 to nearly 75 percent.
- The macro leverage measure (net credit market assets / market value of residential land) rose from 0.64 to 0.93 in absolute value during 1996–2006.
- The gradual introduction and diffusion of securitization and related instruments:
  - First CDOs date back to the early 1980s; the securitization boom began in the mid 1990s with collateralized mortgage obligations (CMOs).
  - Credit default swaps (CDSs) on CMO payments were introduced by the mid 2000s; by end-2007 the market of CDSs alone was worth about $45 trillion (or 3 times U.S. GDP).
  - Synthetic securitization enabled third parties to trade exposures to these income streams.
- Key regulatory changes in the 1990s that affected housing and credit:
  - 1995 New Community Reinvestment Act (strengthened Fannie Mae and Freddie Mac; facilitated mortgage securitization)
  - 1999 Gramm–Leach–Bliley Act (removed the Glass–Steagall separation of banks and non-bank financial firms)
  - 2000 Commodity Futures Modernization Act (stated that derivatives such as CDSs would not be regulated as futures contracts, securities, or lotteries under federal law)
- Financial conditions and lending stance in the run-up:
  - The fraction of banks that tightened credit standards jumped from nearly zero to over 50 percent at the onset of the crisis (Figure 3 evidence).
  - The median downpayment on conventional mortgages rose from 5 to 13 percent (Feb. 16, Wall Street Journal).

### Conceptual mechanism: learning, collateral constraints, and excess returns
- Model environment (representative-agent, incomplete markets):
  - Two assets: a non-state-contingent bond and an asset in fixed supply (land).
  - Borrowing constraint: debt (b_{t+1}/R) cannot be less than −κ_t times market value of land (−κ_t q_t l_{t+1}); κ_t is the regime-dependent leverage fraction.
  - Financial innovation is modeled as introduction of a regime with higher leverage limit κ_h, while the economy can switch to a low-leverage κ_l regime.
- Imperfect information and learning:
  - Agents know κ_h and κ_l but do not know the true regime-switching probabilities F^a.
  - Agents are Bayesian learners who observe realizations of κ without noise and update posteriors on the transition probabilities; beliefs converge in the long run to F^a.
  - Short-run deviation from rational expectations emerges because agents’ posterior beliefs differ from the true transition matrix.
- Anticipated Utility (AU) solution strategy:
  - Two-stage: (i) compute the sequence of posterior densities {f(F^s | κ_t)} by Bayesian updating; (ii) solve a sequence of date-t AU optimization problems conditional on the date-t posterior to generate policy and pricing functions that are then chained to obtain equilibrium dynamics.
  - This approach captures learning about transition probabilities (passive two-point learning) while keeping the optimization computationally tractable.

### Asset pricing implication and role of the collateral constraint
- Forward valuation with shadow value of the credit constraint: land price q_t equals expected discounted sum of future land dividends where stochastic discount factors are adjusted for the Lagrange multiplier μ on the credit constraint (equation (9) in text).
- Excess return on land (exact expression preserved):
  - E_s_t[R^q_{t+1} − R] = ((1−κ_t) μ_t − cov_s_t(λ_{t+1}, R^q_{t+1})) / E_s_t[λ_{t+1}],
  - where R^q_{t+1} ≡ (z_{t+1} g′(l_{t+1}) + q_{t+1})/q_t and λ_{t+1} ≡ β u′(c_{t+1}).
- Channels:
  - Direct effect: (1−κ_t) μ_t raises the excess return when the constraint binds.
  - Indirect effects: the collateral constraint hampers consumption smoothing, reduces cov_s_t(λ_{t+1}, R^q_{t+1}), and lowers E_s_t[λ_{t+1}], all of which affect required returns and asset valuations.
- Interaction between learning and the debt-deflation mechanism:
  - Learning-driven mispricing of risk inflates or deflates collateral values; inflated collateral permits more borrowing, amplifying boom dynamics; deflation tightens collateral, generating sharper contractions.
  - Quantitatively, the model predicts effects on debt and asset prices that are nearly twice as large when both the debt-deflation mechanism and the learning mechanism interact than when either mechanism is removed.

### Quantitative experiment design and baseline calibration (explicit values retained)
- Dating of experiment:
  - Financial innovation begins in 1997Q1; crisis begins in 2007Q1.
  - Optimistic phase: κ_t = κ_h for t = 1, ..., J where J = 40 quarters (1997Q1–2006Q4).
  - Switch to low-leverage: first κ_l realization in 2007Q1; κ_t = κ_l for t = J+1, ..., T, with T = 48 quarters (1997Q1–2008Q4).
- Key calibrated baseline parameter values (preserved exactly):
  - R (real interest rate, annualized) = 2.66 percent.
  - β (discount factor, annualized) = 0.91.
  - σ (risk-aversion coefficient) = 2.0.
  - α (factor share of land in production) = 0.0251.
  - Supply of land L = 1.0.
  - c (consumption-GDP ratio, pre-financial-innovation benchmark) = 0.670.
  - A (lump-sum absorption, exogenous spending residual) = 0.321.
  - Net credit market assets / GDP pre-innovation ( ̄b) = −0.313.
  - κ_l (low-leverage regime) = 0.642 (computed from pre-1997 averages and R).
  - κ_h (high-leverage regime) = 0.926 (set equal to observed 2006Q4 leverage ratio).
  - Markov TFP process approximation: ρ = 0.869, σ_e = 0.00833 (Tauchen–Hussey discretization with 9 states used).
  - Priors for regime-switching probabilities calibrated via beta-binomial counters:
    - Initial symmetric priors chosen so that model matches observed excess return on Fannie Mae 30-year residential MBS spread in 1997Q2 (adjusted, option-adjusted spread) = 47.6 basis points.
    - This calibration yields n_0 = 0.0205 and initial prior Beta(0.0205, 0.0205) (U-shaped with mean 0.5).
  - True transition probabilities (F^a) set so mean duration of high-leverage regime matches estimated credit-boom duration ≈ 7 years:
    - F^a_hh = 0.964, F^a_ll = 0.964 (symmetric true process).
- Learning features and implications of the baseline calibration:
  - With the Beta(0.0205,0.0205) priors, observing an initial long spell of κ_h produces rapid and large optimism: E_t[F^s_hh] moves close to 1 within a few quarters, implying perceived mean durations for κ_h of order 50–1,000 quarters and a much-reduced coefficient of variation of κ.
  - Agents become optimistic about persistence of the high-leverage regime during the optimistic phase, reduce precautionary savings, and increase leverage and land prices above the RE benchmark.
  - Upon observing the first κ_l realization agents update beliefs sharply downward (pessimistic phase), triggering abrupt contractions in credit, land prices, and consumption.

### Quantitative outcomes (selected aggregate implications)
- Contribution of the learning mechanism to observed U.S. pre-crisis dynamics (preserved claims from text):
  - The model accounts for 63 percent of the rise in net household debt during 1997–2006.
  - The model accounts for 44 percent of the rise in residential land prices during 1997–2006.
- Comparative model experiments described:
  - Baseline learning model (BL): includes Bayesian learning, collateral constraint with price-dependent collateral (credit externality), and debt-deflation amplification.
  - Rational expectations (RE) model: retains collateral constraint and amplification but assumes agents know true F^a (no learning friction).
  - Fixed land valuation-learning (FVL) model: retains Bayesian learning but values land in the collateral constraint at a constant long-run average q̄ l, thereby removing the credit externality and the Fisherian deflation channel.
- Key qualitative quantitative findings:
  - Optimistic learning in the presence of a price-dependent collateral constraint leads to pronounced boom in debt, leverage, and land prices that exceed RE predictions; the subsequent realization of a low-leverage regime generates sharp corrections amplified by the collateral channel.
  - Interaction of the learning process with the debt-deflation collateral externality is quantitatively significant: effects on debt and asset prices are nearly twice as large when both mechanisms operate than when either is removed.
- Calibration-based numerical magnitudes summarized (explicit values preserved where provided above):
  - Experiment timing: J = 40 (optimistic spell in quarters), T = 48 (total learning period in quarters).
  - Excess return calibration target: 47.6 basis points (Fannie Mae 30-year MBS spread over T-bill, option-adjusted) → implies n_0 = 0.0205 in priors.
  - κ_l = 0.642, κ_h = 0.926, R = 2.66 percent, β = 0.91, σ = 2.0, α = 0.0251, A = 0.321, c = 0.670,  ̄b = −0.313.
  - True regime persistence: F^a_hh = F^a_ll = 0.964.

### Analytical and policy-relevant implications emphasized by the study
- Financial innovation that expands leverage possibilities can naturally lead to underpricing of risk when agents lack historical data about new instruments and legal/regulatory regimes; Bayesian learning about regime persistence can generate prolonged optimism followed by abrupt pessimistic corrections.
- The collateral constraint linking borrowing to market valuations creates a potent amplification channel: overvaluation of collateral during optimistic phases feeds overborrowing, and subsequent revaluation tightens constraints and magnifies downturns.
- Policy-relevant observations (informed by the model’s mechanism):
  - Regulatory changes that materially alter leverage possibilities create an environment in which history is insufficient to identify regime persistence; policymakers should anticipate learning-driven mispricing and macrofinancial amplification.
  - Monitoring and possibly attenuating rapid expansion in securitization and leverage—particularly when demonstrated performance histories are short—can reduce the buildup of optimism-driven overborrowing.
  - Countercyclical policy tools that limit the direct link between market valuations and collateral-based borrowing (or that strengthen buffers against sharp revaluations) can mitigate the debt-deflation amplification associated with learning reversals.

*Italicized: Source — IMF Working Paper chapter "10. Excess Returns" (excerpt provided).*

### 1.  Ergodic distributions

### 1.  Ergodic distributions

### Key findings about ergodic distributions
- Figure 7 plots “conjectured” ergodic distributions of b for dates 1, 8, 40, 41 and 48 in the BL model and the true ergodic distribution of the RE model.
- The “conjectured” label: the actual ergodic distribution of the BL model equals that of the RE model because in the long-run agents learn F_a and the long-run equilibrium coincides with rational expectations.
- The conjectured ergodic distributions for dates in the learning experiment are agents’ projections of the long-run equilibrium based on their current beliefs about the dynamics of the financial regime.
- Plotting conjectured and RE long-run distributions together illustrates how optimism and pessimism in beliefs affect agents’ willingness to borrow or save.
- Appendix B provides details on the computation of these long-run distributions.

### Specific comparative observations
- Period 1:
  - The mean of bond holdings pre-financial-innovation was -0.31.
  - By period 1 agents conjecture the support of the long-run distribution of bonds will shift sharply to the left (i.e., support higher debt levels).
  - Agents project they will be saving much less than they eventually will in the new stochastic steady state.
  - The RE distribution displays the typical bimodal shape of a two-point regime-switching process with high persistence; agents are correctly assessing the risk of long spells of both κ regimes.
- Period 8:
  - By period 8 agents already conjecture that debt positions in the [-0.54,-0.49] range are most likely long-run equilibria, while in the RE ergodic distribution they have zero probability.
- Period 40 (peak of optimism):
  - Large debt ratios (bond holdings in the interval [-0.59, -0.54]) are never a long-run equilibrium outcome in the RE model, but they take most of the mass of the long-run distribution of bond holdings projected from agents’ period-40 beliefs.
  - As optimism builds, the highest conjectured debt with positive long-run probability rises and the mass assigned to debt levels larger than the largest debt under RE also rises; this process peaks at date 40.
  - During this optimistic phase agents are willing to “overborrow” (take on more debt at the averages of the conjectured ergodic distributions of b) and “undersave” (conjecture they can attain a lower average of b) relative to RE optimality.
- Period 41 onward and period 48 (pessimistic phase):
  - When the first realization of κ_l hits and the pessimistic phase starts, agents reverse behavior: they “underborrow” and “oversave” substantially, peaking by date 48.
  - By period 48 agents have learned shifts to κ_h are possible, so the period-48 conjectured distribution is two-peaked.

### Interpretive summary
- Short sample realizations of κ_h are sufficient for agents to become very optimistic relative to RE transition probabilities, shifting conjectured long-run distributions toward higher debt.
- Subsequent realizations of κ_l reverse beliefs, producing the opposite shift and a two-peaked conjectured distribution once agents recognize possible regime shifts.

---

### 2.  Forecast functions

### Construction and methodology
- Forecast functions are constructed using the sequence of beliefs and decision rules of each AUOP to trace expected values of endogenous variables along the AU recursive equilibrium path for dates 0 to 48.
- The algorithm uses a law of motion for the evolution of the probability of the economy being in each triple (b, z, κ) as time moves from date 0 to date 48.
- The law of motion can be computed for any triple of initial conditions (b, z, κ), but the experiment focuses on the triple approximating the U.S. economy in 1996Q4 (initial conditions at the beginning of date 1).
- Initial conditions used:
  - Date 1 coordinate of initial conditions (b1, z1, κ_h) with b1 = −0.345 (the observed net credit assets as a share of GDP for 1996Q4) and z1 = 1.
  - For each subsequent date, κ is set to the corresponding realization in the κ_T sequence (κ_h for t = 2, ...,40 and κ_l for t = 41, ...,48).
  - Transitions across values of z are computed with the Markov process of z.
  - Transitions across points in the state space of b are governed by the policy function b′_t(b, z, κ) of the date-t AUOP.
- The policy function is time-varying because it varies with each set of beliefs in the sequence [f(F_s|κ_t)]_{48 t=1}.
- The procedure is analogous to standard forecast functions in RE models but conditional on the particular history κ_T used rather than averaging across κ realizations at each date t.

### Forecast function outputs (Figure 8)
- Figure 8 plots forecast functions for:
  - bond holdings as a share of output (b′/y),
  - consumption,
  - the price of land,
  - the savings rate (GSF/y),
  as percent deviations from their long-run means in the BL, RE and FVL models.
- Notes on mean comparisons:
  - Long-run means in the learning and RE models are identical because their ergodic distributions are the same.
  - The solid (blue) lines correspond to the BL model, dashed (green) lines to the FVL model, and dotted (red) lines to the RE model.
  - Even the RE model generates dynamics in this exercise because the initial condition b1 is not the long-run mean of the new financial regime with stochastic κ, and because a particular time series of realizations of κ is used rather than averaging.

### Comparative dynamics for bonds
- Panel (a) forecast functions for bonds show:
  - During the optimistic phase agents consistently borrow more in the BL model than in both the RE and the FVL models.
  - In the first two periods after financial innovation is introduced, the three models predict similar debt dynamics.
  - After that, the optimism and the debt-deflation feedback loop in the BL model produce a much larger decline in bond holdings, while bond dynamics in the RE and FVL models remain similar.
  - Bond holdings as a share of output decline by as much as 21 percentage points below the long-run average at the peak of optimism of the BL model in period 40.

*Source: _wp10164 - 1.  Ergodic distributions*

### 40. These dynamics are in line with the downward trend in net credit market assets observed in

### _wp10164 - 40. These dynamics are in line with the downward trend in net credit market assets observed in

### Learning, debt-deflation, and alternative model mechanisms
- The combination of the learning friction and the debt-deflation channel (BL model) delivers a much stronger decline in assets than alternatives that retain only one mechanism.
- RE model: no buildup of optimism to push for overborrowing; no endogenous feedback from higher land prices into higher collateral and thus higher borrowing ability.
- FVL model: no endogenous Fisherian deflation feedback from higher land prices into collateral; price used for collateral is fixed.
- The switch to the pessimistic phase in period 41 produces a large correction in bond holdings:
  - Bond holdings bounce back about 54 percentage points in the BL model at the date-41 switch.
  - Debt reversal in the RE case is about half the size of that in the learning model.
  - FVL model yields the smallest correction (isolates switch to pessimistic beliefs without Fisherian amplification).
- The large κ shock at date 41 behaves like an unexpected shock because by date 40 agents believe state κh is almost absorbent (E40[Fs hh] very close to 1), amplifying the equilibrium adjustment despite CRRA preferences and incomplete markets.

### Forecast functions, asset demand, and land prices (Figure 8 discussion)
- Overborrowing in the optimistic phase (BL) increases demand for risky asset and raises land prices substantially (Panel (b)):
  - BL: significant land price boom.
  - RE: price of land declines slightly relative to pre-financial-innovation price at date 0 because stochastic κ increases mean variance—variance effect can dominate when gap between κh and κl is wide.
  - FVL: larger asset price boom during optimistic phase and smaller crash because fixed collateral valuation rules out Fisherian deflation and provides a limited asset-price-for-collateral guarantee (fixed only for collateral valuation, not market trade price).
- The FVL model produces a larger land price increase in optimistic phase but this does not feed back into a comparably large debt expansion due to fixed collateral valuation.

### Consumption and gross saving flow dynamics (Panels (c), (d), (e), (f))
- Consumption tilting effect from switching κl → κh in perfect-foresight: initial rise above pre-financial-innovation level, then monotonic drop to a new lower steady state because higher steady-state debt lowers steady-state consumption when non-financial wealth invariant.
- In stochastic and learning models:
  - Precautionary savings weaken the consumption tilting effect (smaller decrease in bond holdings and smaller drop in consumption).
  - For first 40 periods, consumption dynamics similar across BL, RE, FVL because consumption converges quickly to new long-run average:
    - Overconsumption in BL relative to RE and FVL between period 3 and 10 due to larger increase in debt in BL.
    - After period 10, consumption tilting dominates until date 41 when κ switches to κl.
- Date 41 switch:
  - BL with Fisherian deflation produces dramatic collapse in consumption (consistent with literature noting possible equilibrium drops around -80 percent in deflationary equilibria; learning friction supports large short-run AU equilibria).
  - RE and FVL produce large consumption declines as well but significantly smaller than BL:
    - RE: precautionary savings and lack of overborrowing attenuate decline.
    - FVL: absence of Fisherian deflation yields smallest consumption drop.

### Asset price dynamics, interest premia, and risk compensation (Figure 9 discussion)
- Interest rate premium (difference between stochastic intertemporal MRS and real interest rate R; equals μt/(u′(ct)−μt) using condition (12)):
  - In optimistic phase the collateral constraint becomes nonbinding initially then binds after several periods:
    - BL: constraint begins to bind after period 5; converges to about 5.5 percent at peak of optimistic phase.
    - FVL: larger premium up to 7 percent in optimistic phase.
    - RE: premium just above 2 percent in optimistic phase.
  - At switch to pessimistic phase (date 41) there is a large jump in interest premia; ordering of correction largest to smallest: BL, RE, FVL.
  - After date 41:
    - Constraint becomes nonbinding for 4 periods in BL and 1 period in RE.
    - Interest premia then positive and hover around 7 percent in FVL and increase to reach 8.9 percent (BL) and 8.2 percent (RE) by end of experiment.
- Period-41 reported values (from notes to Figure 9):
  - Interest premium in period 41: 38.03 (BL), 1.99 (RE), 0.81 (FVL).
  - Sharpe ratio in period 41: 134.71 (BL), 48.59 (RE), 5.16 (FVL).
  - Expected excess return in period 41: 13.94 (BL), 0.73 (RE), 0.30 (FVL).
- Sharpe ratio and components:
  - Sharpe ratio St ≡ Es t [Rq t+1 − R] / σs t(Rq t+1) rises in all models during optimistic phase, with larger and more gradual increases in BL and FVL than RE.
  - In BL and FVL the rise in Sharpe ratio mainly driven by gradual declines in σs t(Rq t+1) while Es t [Rq t+1 − R] remains largely stable — perceived riskiness of land declines.
  - Up to period 40, BL and FVL Sharpe ratios are similar, but BL exhibits higher mean excess returns and higher variability of returns than FVL.
- Price of risk st = σs t(Mt+1 t) / Es t[Mt+1 t]:
  - In optimistic phase:
    - RE and BL: price of risk increases initially; RE settles about 0.22, BL rises to about 0.5 by period 40.
    - FVL: price of risk lower initially and falls slightly during optimistic phase (implicit land-price guarantee reduces price of risk).
  - At period 41 both σs t(Mt+1 t) and Es t[Mt+1 t] fall sharply; larger fall in σ makes price of risk fall.
- Direct effect of collateral constraint on expected excess returns (Panel (e)):
  - FVL scenario: direct effect > 90 percent of expected excess returns by period 40.
  - BL: direct effect rises gradually to about 30 percent by period 40.
  - RE: contribution stable at about 10 percent from periods 5 to 40.
  - When switch to κl occurs, direct effect grows very large in all models as constraint becomes binding; thereafter it remains large in FVL and falls back to zero in BL and RE as debt adjusts, then rises again as constraint rebinds.
- Early-stage drivers of increasing excess returns:
  - When collateral constraint does not bind, increases in mean excess returns in RE and BL are driven solely by covs t(Mt+1 t, Rq t+1) becoming more negative (i.e., declining covariance), reducing perceived riskiness.

### Turning points and quantitative comparisons with data (Table 2 discussion)
- Table 2 reports average changes at two turning points:
  - Peak of optimism at t = 40 relative to pre-financial-innovation initial conditions (period 0 = 1996Q4 observation).
  - Financial crisis: t = 48 relative to peak t = 40.
- Reported values (differences in levels of b/y and qL/y projected by forecast functions; qL/y = aggregate market value of residential land divided by output):
  - Peak of Optimism:
    - Data: E[(b/y)40 − (b/y)0] = -0.355
    - RE: -0.083
    - FVL: -0.089
    - BL: -0.223
    - Data: E[(qL/y)40 − (qL/y)0] = 0.280
    - RE: -0.025
    - FVL: 0.305
    - BL: 0.122
  - Financial Crisis:
    - Data: E[(b/y)48 − (b/y)40] = 0.023
    - RE: 0.122
    - FVL: 0.133
    - BL: 0.254
    - Data: E[(qL/y)48 − (qL/y)40] = -0.149
    - RE: 0.013
    - FVL: -0.305
    - BL: -0.121
  - Note: Data column top panel is difference between 2006Q4 and 1996Q4; bottom panel difference between 2008Q4 and 2006Q4.
- Key quantitative findings:
  - BL model explains a significant part of pre-crisis increases in debt and land prices.
  - BL generates significantly higher debt in optimistic phase than RE or FVL and much larger land price increase than RE.
  - BL explains 63 percent of increase in household debt observed in data (model b/y falls by almost 23 percentage points v. 36 percentage points in the data).
  - BL decline in bond holdings is about 14 percentage points of GDP larger than in RE or FVL.
  - BL accounts for about 44 percent of land price boom in data (model q increases almost 13 percentage points v. 28 percentage points in data).
- Crisis-period comparisons:
  - BL model generates a large debt reversal of about 25 percentage points (between t = 40 and t = 48 the BL model reversal is large; data correction was only 2.3 percentage points).
  - Model overestimates debt reversal partly because model bonds are one-period while household debt in data has much higher average maturity (e.g., 30-year mortgages), so real-world reversal is spread over time.
  - By 2Q 2010 reversal in household debt ratio reached about 10 percentage points of GDP in data (as noted in text).
  - BL matches observed decline in land prices during financial crisis reasonably well: model decline 12.1 percentage points v. data decline 14.9 percentage points.
  - FVL model produces a larger price decline than data (about twice as large).
  - RE model produces small price change and in the opposite direction to BL and FVL.

*Source: Excerpt from _wp10164 - 40. These dynamics are in line with the downward trend in net credit market assets observed in (PDF chapter).*

### 4.  Projected Excess Returns on Land

### 4.  Projected Excess Returns on Land

### Overview of projections and setup
- Expected excess returns are computed t+j-periods ahead for up to 50 periods, using beliefs and decision rules as of initial dates t = 1, 40, and 41 (panels (a), (b), and (c) respectively).
- For each scenario the initial state is set so that:
  - b is at the mean bond holdings predicted by the forecast function in Figure 8 for the corresponding date,
  - κ is set to its corresponding value in the history κ_t,
  - TFP is set to its mean value (z = 1).
- Agents form expectations conditional on these initial states and the date-specific equilibrium pricing function.

### Key comparative findings across models (RE, BL, FVL)
- Date 1 projections (panel (a)):
  - RE model excess returns slightly exceed those of BL up to the 10th period; after period 10 the ordering reverses and BL projects slightly higher returns.
  - FVL model yields expected excess returns significantly below both RE and BL models.
  - Implication: date-1 land price is slightly lower in RE than in BL (RE agents expect relatively higher excess land returns in first 10 periods). FVL’s lower excess returns are consistent with a sharply higher date-1 land price.
  - Explanation: removal of the debt-deflation channel in FVL weakens direct and indirect effects of the collateral constraint on excess returns (see Equation (10)).
- Date 40 projections (panel (b)):
  - As agents reach the optimistic-phase peak after observing κ_h for 40 periods, expected excess returns ahead of date 40 are significantly lower than those predicted as of date 1 in BL and FVL.
  - BL and FVL produce sharply higher land prices at date 40 than at date 1.
  - Projected returns in BL become significantly smaller than in RE; FVL predicts even smaller excess returns than RE and BL.
  - Reason: FVL beliefs are as optimistic as the learning model but, without the debt-deflation mechanism, risk premia on holding land are reduced.
- Date 41 projections (panel (c)):
  - At the switch to the low-leverage regime (κ switches to κ_l), ordering of projected excess returns across RE and BL reverses.
  - Projected excess returns for period 42 are very high in all three models because of the strong direct effect of the borrowing constraint tightening sharply as κ switches. The direct effect includes:
    - An exogenous effect from the switch to κ_l, and
    - An endogenous effect via the shadow value of the borrowing constraint (for the excess return at t = 42 expected as of t = 41, the direct effect in the right-hand side of (10) is given by (1−κ_l)μ_41).
  - The direct effect is strongest in BL (learning + Fisherian deflation), next in RE (retains Fisherian deflation), and weakest in FVL.
  - Interest rate premia ordering in period 41 mirrors this size and ordering.

### Dynamics after the switch and adjustment phase
- Immediately after the severe tightening and abrupt debt adjustment:
  - In BL and RE, the borrowing constraint is projected not to bind for a couple of periods, then to bind again as debt is rebuilt.
  - In FVL the constraint is projected to remain binding, but debt adjustment reduces tightness sharply and thus projected returns decline.
- Beyond the first 10-period adjustment phase:
  - Projected returns in BL exceed those of RE and FVL.
  - Projected returns in FVL are sharply lower than in the other two models.
  - Consistency: at date 41 the value of q is highest in FVL, followed by RE, with BL price sharply lower than the other two (legend values shown in Figure 10: BL, q = 0.045; FVL, q = 0.480; RE, q = 0.294 in panel (a); BL, q = 0.598; FVL, q = 0.781; RE, q = 0.451 in panel (b); BL, q = 0.478; FVL, q = 0.662; RE, q = 0.465 in panel (c)).
- Short-run ordering nuance:
  - During periods 2 to 7 ahead of date 41, projected excess returns of RE exceed those of BL.
  - Interpretation: pessimistic expectations in BL result in slower debt buildup, so the collateral constraint is expected to start binding one period later than under RE and to bind with lower shadow values (lower μ’s) than under RE until period 10.
  - Long run: as of date 41 beliefs still favor overborrowing relative to rational expectations; agents project the borrowing constraint will eventually become more binding in BL than in RE, so land returns are projected to converge to a higher level under BL.

### Specific numeric and illustrative values reported
- Projections computed for 50 periods ahead of initial dates t = 1, 40, and 41.
- One-period-ahead expected excess return in period 41 (bottom panel) is reported as:
  - 1418 percent in baseline,
  - 74 percent in RE,
  - 30 percent in FVL.
- Figure 10 legend q values (per panel):
  - Panel (a) Period 1: BL, q = 0.045; FVL, q = 0.480; RE, q = 0.294.
  - Panel (b) Period 40: BL, q = 0.598; FVL, q = 0.781; RE, q = 0.451.
  - Panel (c) Period 41: BL, q = 0.478; FVL, q = 0.662; RE, q = 0.465.
- TFP mean value used in conditioning: z = 1.

*Source: IMF working paper section "4.  Projected Excess Returns on Land"*

### 5.  Sensitivity Analysis

### 5. Sensitivity Analysis

### Overview
- The sensitivity analysis focuses on turning point effects and examines:
  - Changes in initial priors (Table 3 / Scenarios (1)–(4)).
  - Changes in κh, κl, β, and R (Table 4 / Scenarios (1)–(4)).
- General finding: parameterizations that generate larger booms during the optimistic phase also generate larger busts during the financial crisis.

### Sensitivity to Initial Priors (Scenarios summarized from Table 3 and text)
- Baseline logic: the size of booms and busts depends critically on agents’ initial knowledge about the transition and continuation probabilities of κh. Having never observed κh, and particularly not having seen a transition from κh to κl, is crucial to generate sizable booms and busts.

- Scenario (1): uniform priors (n0 = 1)
  - Debt dynamics qualitatively similar to BL but smaller debt buildup: 9 percentage points v. 22 in BL.
  - Land price movements: price of land falls to a level about 2 percentage points lower in period 40 than in period 1 (vs. nearly 12 percentage points increase in BL).
  - Land price trajectory during optimistic phase is u-shaped (not monotonically increasing as in BL) because beliefs about persistence of κh remain lower than in BL; higher variance dominates higher mean early on.
  - Financial crisis effects much weaker: debt adjusts by 14 percentage points instead of 25; price of land falls only by 0.6 of a percentage point, instead of nearly 12 percentage points.

- Scenario (2): lower initial counters (n0 = 0.01)
  - Priors concentrate more mass near 0 and 1 than BL (BL calibrated to n0 = 0.0205).
  - Agents turn more optimistic after observing the first κh, leading to larger debt and land-price booms: debt buildup about 24 percentage points; land price boom reaches 14 percentage points.
  - Pessimistic phase produces larger correction in debt and larger drop in land prices.

- Scenario (3): set nhh0 = nll0 = 0.54 while keeping nhl0 and nlh0 at BL values (0.0205)
  - Initial means equal to true persistence parameters of κ regimes, but agents are still uncertain and update beliefs.
  - Results very similar to BL model (initial differences in optimism shrink quickly).

- Scenario (4): informative κl realizations in pre-innovation period (nll0 = 68, nhh0 = 1)
  - Priors imply Es0[Fll] = 0.985 v. Es0[Fll] = 0.5 in BL.
  - During early optimistic phase, conjectured ergodic distributions of bonds are double-peaked with large mass around mean conditional on κl → unconditional mean debt smaller than BL during optimistic phase.
  - During pessimistic phase, higher debt levels than BL because agents are less pessimistic after having observed one κl→κh transition.
  - Despite these differences, turning-point dynamics in Table 3 do not differ much from BL because forecast functions for turning points use histories conditional on κh during optimistic phase and κl during pessimistic phase.

- Quantitative entries from Table 3 (as reported):
  - E[(b/y)40 − (b/y)0]: BL −0.223; n0 = 1 −0.087; n0 = 0.01 −0.243; nhh0 = 0.54 −0.222; nll0 = 68 −0.221
  - E[(ql/y)40 − (ql/y)0]: BL 0.122; n0 = 1 −0.020; n0 = 0.01 0.142; nhh0 = 0.54 0.121; nll0 = 68 0.121
  - E[(b/y)48 − (b/y)40]: BL 0.254; n0 = 1 0.137; n0 = 0.01 0.274; nhh0 = 0.54 0.253; nll0 = 68 0.254
  - E[(ql/y)48 − (ql/y)40]: BL −0.121; n0 = 1 −0.006; n0 = 0.01 −0.141; nhh0 = 0.54 −0.120; nll0 = 68 −0.123

- Key interpretation:
  - Initial priors calibrated to data (the BL exercise matching observed excess return on 1997:Q1 Fannie Mae RMBS) are important to discipline the magnitude of model effects.
  - Priors can be set arbitrarily to generate larger or smaller effects; a data-based criterion helps assess relevance.

### Sensitivity to Other Parameters (Table 4: R, κl, κh, β)
- Experiment framing:
  - Scenario (1) changes R to the average interest rate for 1997-2008 in the data: 0.98 percent (as opposed to 2.66 percent used in BL).
  - Scenario (2) increases κl to 0.75.
  - Scenario (3) reduces κh to 0.8.
  - Scenario (4) increases β to 0.95 (BL β = 0.91).

- Quantitative entries from Table 4 (as reported):
  - E[(b/y)40 − (b/y)0]: BL −0.223; R = 1.0098 −0.248; κl = 0.75 −0.240; κh = 0.8 −0.143; β = 0.95 −0.237
  - E[(ql/y)40 − (ql/y)0]: BL 0.122; R = 1.0098 0.158; κl = 0.75 0.140; κh = 0.8 0.120; β = 0.95 0.140
  - E[(b/y)48 − (b/y)40]: BL 0.254; R = 1.0098 0.208; κl = 0.75 0.157; κh = 0.8 0.175; β = 0.95 0.175
  - E[(ql/y)48 − (ql/y)40]: BL −0.121; R = 1.0098 −0.038; κl = 0.75 −0.059; κh = 0.8 −0.120; β = 0.95 0.098

- Scenario (1): lower R (0.98 percent vs. BL 2.66 percent)
  - Long-run rational expectations means change: E[b/y] goes down from −0.36 in the RE and BL to −0.41; land price increases from 0.45 to 0.53.
  - Lower R generates larger increases in debt and asset prices during the optimistic phase (asset prices need to rise when R falls to equate expected returns).
  - Crisis correction in debt is about 5 percentage points smaller; decline in price of land is significantly smaller.

- Scenario (2): higher κl (κl = 0.75)
  - Slightly larger booms in debt and land prices because low-leverage regime is not as low → higher mean and lower variance of κ.
  - Reversals in debt and land prices are smaller because higher κl allows more borrowing during pessimistic phase.

- Scenario (3): lower κh (κh = 0.8)
  - Tighter credit constraint in high-leverage regime reduces size of debt buildup and of debt reversal in crisis.
  - Land price boom and crash change only marginally.

- Scenario (4): higher β (β = 0.95 v. 0.91 in BL)
  - Higher discount factor supports higher asset prices and slightly larger debt and land-price booms in the optimistic phase.
  - Financial crisis effects with higher β are weaker, particularly for asset prices which recover quickly and reach levels higher than pre-crisis by period 48.

### Key quantitative magnitudes and mechanisms (from analysis and conclusion)
- The calibrated model implies:
  - The interaction of learning frictions and a Fisherian collateral constraint generates substantial overborrowing.
  - Overborrowing accounts for almost two-thirds of the increase in net debt of U.S. households between 1997 and 2006.
  - Overborrowing accounts for about two-fifths of the boom in residential land prices between 1997 and 2006.
  - Overborrowing from learning alone (without the credit externality) is about one half of that produced when both learning and the credit externality operate.

### Policy implications and recommendations (from Conclusion)
- Financial innovation carries inherent uncertainty: when a truly brand-new financial regime starts, the true riskiness cannot be correctly evaluated due to lack of data.
  - Recommendation: close supervision of financial intermediaries in the early stages of financial innovation is critical.
- Limitations of taxes or capital requirements intended to manage “macroprudential risk”:
  - Taxes or capital requirements can address overborrowing due to the credit externality (pecuniary externality), but they cannot correct overborrowing driven by optimistic beliefs if policymakers are as uninformed as households.
  - Nonetheless, taxes/capital requirements can still help: the overborrowing effect of learning without the credit externality is about one half of that when both channels are present.
- Ongoing regulatory reforms (e.g., Basle III, Dodd-Frank) are themselves radical innovations in capital markets:
  - Agents must evaluate the riskiness of the new financial environment under imperfect information, implying risk that multi-year pessimism can hamper recovery.

*Source: 5. Sensitivity Analysis (excerpt) from the provided IMF PDF content.*

### References

### _wp10164 - References

### References (topics and literature covered)
- Empirical and theoretical work on house prices, current account, and financial crises: Adam, Klaus, Pei Kuang, and Albert Marcet, 2011; Bernanke, Ben, 2005; Bianchi, Javier, 2009; Boz, Emine, 2009; Durdu, Ceyhun Bora, Enrique G. Mendoza, and Marco Terrones, 2009; Mendoza, Enrique G., 2010; Mendoza, Enrique G., and Vincenzo Quadrini, 2010; Mendoza, Enrique G., Vincenzo Quadrini, and Jos`e-V`ıctor R`ıos-Rull, 2009; Mendoza, Enrique G., and Katherine A. Smith, 2006; Mendoza, Enrique G., and Marco Terrones, 2008.
- Learning, forecasting, and excess volatility in asset markets: Barsky and De Long, 1993; Blanchard, L’Huillier, and Lorenzoni, 2008; Cogley and Sargent, 2008a; Cogley and Sargent, 2008b; Edge, Laubach, and Williams, 2007; Evans and Honkapohja, 1999; Timmermann, 1993 and 1996; Van Nieuwerburgh and Veldkamp, 2006; Zeira, 1999.
- Housing, mortgage innovation, and debt-deflation mechanisms: Campbell and Hercowitz, 2009; Corbae and Quintin, 2009; Davis and Heathcote, 2007; Favilukis, Ludvigson, and Van Nieuwerburgh, 2010; Fisher, 1933; Jermann and Quadrini, 2006; Goetzmann and Newman, 2010; Piazzessi and Schneider, 2009.
- Financial intermediation, leverage, contagion, and externalities: Gertler and Kiyotaki, 2010; Korinek, 2008; Lorenzoni, 2009; Reinhart and Rogoff, 2009; Warnock and Warnock, 2006; Uribe, 2006.
- Numerical methods and approximation techniques: Tauchen and Hussey, 1991; Bulkley and Tonks, 1989; Lansing, 2009.

### Appendix A — Solution Method (AUOP recursive solution)
- Solve each AUOP using an Euler equation method combining price and policy function iterations with the land pricing equation and general equilibrium conditions (12)-(15).
- Algorithm steps:
  - 1. Define a history of realizations κ_T and calculate the sequence of posteriors {f(F_s | κ_t)}_{t=1}^T.
  - 2. Take κ_1 and the date-1 posterior f(F_s | κ_1) and compute E_s1[κ′ | κ] ≡ [ E_1[F_s^hh]   1−E_1[F_s^hh] / 1−E_1[F_s^ll] E_1[F_s^ll] ].
  - 3. Using E_s1[κ′ | κ] and a guess for the land pricing function q_1(b_t, z_t, κ_t), solve for the date-1 equilibrium conditions using a policy function iteration algorithm.
  - 4. Use resulting policy functions [b′_1(b, z, κ), c_1(b, z, κ), μ_1(b, z, κ)] and the asset pricing equation (11) to compute a new pricing function ˆq_1(b_t, z_t, κ_t).
  - 5. Compare ˆq_1(b_t, z_t, κ_t) and q_1(b_t, z_t, κ_t). If they satisfy a convergence criterion then the decision rules [b′_1(b, z, κ), c_1(b, z, κ), μ_1(b, z, κ)] and q_1(b_t, z_t, κ_t) are the date-1 AUOP solutions. If not, construct a new guess using a Gauss-Siedel rule and return to Step 3.
  - 6. Move to date-2 with history κ_2 and posterior f(F_s | κ_2) from Step 1. Compute the Markov transition matrix E_s2[κ′ | κ], return to Step 3, and repeat for each date-history κ_t and posterior f(F_s | κ_t) for t = 1, ... T solving each time the corresponding date-t AUOP.
- Implementation notes on passive Bayesian learning:
  - 1. The solutions to each date-t AUOP are not functionally related; solving date-t does not require the solution for any other date. The model can be solved by solving each date-t AUOP independently. Speed-up: when ||f(F_s | κ_{t+j}) − f(F_s | κ_t)|| is small enough under some metric, use date-t AUOP initial guesses for date-t+j AUOP.
  - 2. If j ≤ T is large enough for f(F_s | κ_{t+j}) to converge to F_a (for some convergence criterion), the solutions for all dates t ≥ j collapse to a standard recursive RE equilibrium using the true Markov process F_a.
  - 3. The full intertemporal equilibrium from dates 1 to T is obtained by chaining solutions of each date-t AUOP. The authors consider only one history κ_T intended to match observed financial regimes between 1997 and 2007; an alternative would be to generate I potential histories [κ_T^i]_{i=1}^I using the true Markov process F_a, solve for each, and average across solutions.

### Appendix B — Computation of Ergodic Distributions, Forecast Functions, Excess Returns
- B.1 Ergodic Distribution and Forecast Functions under Rational Expectations:
  - Define date-t probability distribution over bonds, productivity and collateral coefficients in the RE model as λ_t(b, z, κ).
  - Law of motion:
    - λ_{t+1}(b′, z′, κ′) = ∑_z ∑_κ ∑_{ {b: b′ = g(b,z,κ)} } λ_t(b, z, κ) π(z′ | z) p(κ′ | κ)
    - where g(b, z, κ) is the policy function, π(z′ | z) is productivity Markov transition, and p(κ′ | κ) is true Markov transition of κ.
  - The unconditional limiting distribution λ(b, z, κ) is the fixed point of the law of motion. The algorithm iterates the law of motion until λ_t and λ_{t+1} satisfy a convergence criterion.
  - Forecast functions: averages of endogenous variables at each date t using distribution λ_t(b, z, κ), starting from initial condition (b_0, z_0, κ_0) with λ_0(b_0, z_0, κ_0) = 1. Forecast functions converge to unconditional long-run averages computed with the ergodic distribution (if unique and invariant).
  - Conditional date-t probability distribution over κ_i for i = h, l:
    - ˜λ_t(b, z | κ_i) = λ_t(b, z, κ_i) / ∑_b ∑_z λ_t(b, z, κ_i)
  - Conditional forecast functions are averages computed using ˜λ_t(b, z | κ_i) and converge to corresponding long-run conditional averages ˜λ(b, z | κ_i).

- B.2 AU Forecast Functions in the Learning Model:
  - Under learning, beliefs about transition probability matrix of κ evolve; RE definitions do not apply.
  - Forecast functions in the learning model average only over productivity shocks and use date-specific decision rules produced by the corresponding beliefs and AUOP.
  - Formulae for bonds forecast (given (b_0, z_0, κ_0) and sequence of realizations [κ_t]_{t=0}^T and beliefs [f(F_s | κ_t)]_{t=0}^T):
    - ˆb_1 = E[b_1 | (b_0, z_0, κ_0), f(F_s | κ_0)] = h_0(b_0, z_0, κ_0; f(F_s | κ_0))
    - ˆb_2 = E[b_2 | b_0, f(F_s | κ_1)] = ∑_{z_1} ∑_{ {b_1: b_2 = h_1(b_1, z_1, κ_1)} } π(z_1 | z_0) h_1(b_1, z_1, κ_1; f(F_s | κ_1))
    - ˆb_3 = E[b_3 | b_0, f(F_s | κ_2)] = ∑_{z_2} ∑_{ {b_2: b_3 = h_2(b_2, z_2, κ_2)} } π(z_2 | z_0) h_2(b_2, z_2, κ_2; f(F_s | κ_2))
    - ...
    - ˆb_{T+1} = E[b_{T+1} | b_0, f(F_s | κ_T)] = ∑_{z_T} ∑_{ {b_T: b_{T+1} = h_T(b_T, z_T, κ_T)} } π(z_T | z_0) h_T(b_T, z_T, κ_T; f(F_s | κ_T))
  - π(z_t | z_0) = π(z_t | z_{t−1}) π(z_{t−1} | z_{t−2}) ... π(z_1 | z_0)
  - h_t(b_t, z_t, κ_t; f(F_s | κ_t)) is the optimal bond decision rule from the date-t AUOP using date-t beliefs.
  - Equivalent RE forecast objects use g(b, z, κ) policy function:
    - ˜b_1 = E[b_1 | (b_0, z_0, κ_0)] = g(b_0, z_0, κ_0)
    - ˜b_2 = E[b_2 | b_0] = ∑_{z_1} ∑_{ {b_1: b_2 = g(b_1, z_1, κ_1)} } π(z_1 | z_0) g(b_1, z_1, κ_1)
    - etc.
  - Define χ_t(b, z) as the distribution of TFP shocks and bond holdings at date t in the learning model. Law of motion (given κ_T):
    - χ_{t+1}(b′, z′) = ∑_b ∑_z χ_t(b, z) π(z′ | z) I_t(b′, b, z, κ_t)
    - where I_t(b′, b, z, κ_t) = 1 ↔ b′ = h_t(b, z, κ_t; f(F_s | κ_t)) and zero otherwise.
  - Example initial condition χ_0(b_0, z_0) = 1; χ_1(h_0(b_0, z_0, κ_0; f(F_s | κ_0)), z′) = π(z′ | z_0) for each z′ and zero otherwise.
  - Expected bonds at date 2:
    - ˆb_2 = E[b_2 | b_0, f(F_s | κ_1)] = ∑_b ∑_z χ_1(b, z) h_1(b, z, κ_1; f(F_s | κ_1))
  - Alternative compact law of motion:
    - χ_{t+1}(b′, z′) = ∑_z ∑_{ {b: b′ = h_t(b, z, κ_t; f(F_s | κ_t))} } χ_t(b, z) π(z′ | z)

- B.3 Expected Returns j Periods Ahead of Date t:
  - Choose initial triple (b_t, z_t, κ_t) with initial bond holdings b_t = ˆb_t (mean bond holdings at period t from forecast functions). Set z_t = 1. κ_t as in forecast function calculations for period t.
  - Expected returns for datet + 1 + j use numerator [ q_t(b_{t+j+1}, z_{t+j+1}, κ_{t+j+1}) + d(z_{t+j+1}) ] and denominator q_t(b_{t+j}, z_{t+j}, κ_{t+j}), all projected as of datet.
  - Procedure:
    - Step 1: Calculate probability tree of possible states conditional on (b_t, z_t, κ_t) up to J periods ahead. Events combine TFP and κ shocks.
    - Step 2: Construct sequence E_s_t[R^q_{t+1+j}] for j = 0,1, ..., J.
    - Cross-check: recover asset price q_t(b_t, z_t, κ_t) using sequence E_s_t(1 / E_s_{t+j}[R^q_{t+1+j}]) to recalculate date-t price as present discounted value of dividends discounted by expected returns.
  - Initial probability mass:
    - λ^t_t(b_t, z_t, κ_t) = 1 for j = 0.
  - Forward evolution for j = 0,1, ..., J:
    - λ^t_{t+j+1}(b_{t+j+1}, z_{t+j+1}, κ_{t+j+1}) = ∑_{z_{t+j+1}} ∑_{κ_{t+j+1}} ∑_{b_{t+j+1} ∈ H_{t+j+1}} λ^t_{t+j}(b_{t+j}, z_{t+j}, κ_{t+j}) π(z_{t+j+1} | z_{t+j}) E_s_t(κ_{t+j+1} | κ_{t+j})
    - where H_{t+j+1} = { b_{t+j+1} : b_{t+j+1} = h_t(b_{t+j}, z_{t+j}, κ_{t+j} ; f(F_s | κ_t)) }.
    - The superscript t denotes that these probabilities are conditional on date t initial conditions.
  - Expected returns conditional on date t:
    - E_s_t[R^q_{t+1+j}] = ∑_{z_{t+j+1}} ∑_{κ_{t+j+1}} ∑_{b_{t+j+1} ∈ H_{t+j+1}} ∑_{z_{t+j}} ∑_{κ_{t+j}} λ^j_{t+j}(b_{t+j}, z_{t+j}, κ_{t+j}) π(z_{t+j+1} | z_{t+j}) E_s_t(κ_{t+j+1} | κ_{t+j}) × [ q_t(b_{t+j+1}, z_{t+j+1}, κ_{t+j+1}) + d(z_{t+j+1}) ] / q_t(b_{t+j}, z_{t+j}, κ_{t+j})
    - Note: d(z) = z g′(l). E_s_t[R^q_{t+1+j}] is in fact E_s_t[R^q_{t+1+j}](b_{t+j}, z_{t+j}, κ_{t+j}); one-period-ahead expected returns depend on date-j triple (b_{t+j}, z_{t+j}, κ_{t+j}).
  - Recalculation of q_t(b_t, z_t, κ_t) as present discounted sum of future dividends (using Equation 11):
    - q_t(b_t, z_t, κ_t) = ∑_{j=0}^J ∑_{z_{t+j+1}} ∑_{κ_{t+j+1}} ∑_{b_{t+j} ∈ H_{j+1}} ∑_{z_{t+j}} ∑_{κ_{t+j}} λ^j_{t+j}(b_{t+j}, z_{t+j}, κ_{t+j}) π(z_{t+j+1} | z_{t+j}) E_s_t(κ_{t+j+1} | κ_{t+j}) × ( ∏_{i=0}^j ( 1 / E_s_t[R^q_{t+1+i}] ) ) d(z_{t+j+1})
    - Dividends at date t+j+1 are discounted by the product of reciprocals of one-period-ahead expected returns up to that date. The probability tree from Step 1 supplies the needed probabilities.

*Source: _wp10164 - References*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2010/_wp10164.pdf_
