A Learning Model of Financial Instability
What this paper finds — and why it matters
Layer 1: Overview
Research question and motivation: Williams asks whether the recurrent boom-bust dynamics of Minsky’s financial instability hypothesis — “periods of stability lead to periods of instability” — can arise endogenously from a tractable rational-agent model in which investors learn about asset returns. This matters because standard rational-expectations asset-pricing models cannot generate the high, volatile price-dividend ratios, sizeable risk premia, and recurrent crashes seen in data, and because Minsky’s narrative has long lacked a clean formal mechanism. The paper’s main contribution is theoretical (a new instability/limit-cycle result for adaptive learning), with a secondary quantitative exercise.
Model setup: A small-open-economy variant of the Lucas (1978) consumption-based asset-pricing model studied under learning by Adam, Marcet and Nicolini (2016). A representative agent with power utility (risk aversion gamma, discount factor beta) can borrow/lend at a fixed risk-free gross return R and holds a unit supply of stock paying an i.i.d.-growth dividend (log dividend growth = d + sigma*W, with centered binomial shocks W in {-1,1}). Adding the risk-free asset creates a portfolio problem and endogenous debt dynamics (the net asset position omega), which the closed-economy literature lacks. Agents wrongly believe log returns are i.i.d. binomial with mean m and standard deviation s, and update (m, s^2) by constant-gain recursive least squares with gain epsilon (the weight on new information). A borrowing/leverage constraint (0 <= v <= vbar on the stock portfolio share) ensures equilibrium exists. The self-confirming equilibrium (SCE) has (m,s)=(mu,sigma), v=1, omega=1, and a constant price-dividend ratio.
Mechanism: The pricing function is extremely steep near v=1; the derivative at the SCE is delta’(1)=delta*(1+delta*), so with a mean P/D near 29 a 1-percentage-point fall in v (to 0.99) implies roughly a 30% drop in P/D (to ~20.3). Tranquil periods lower volatility estimates, raising v and prices; once heavily invested, the economy is fragile. Booms end via two mechanisms: binding leverage constraints (rare in the calibration, driving only one crash in the long simulation) and — the novel and dominant channel — a rapid boom raising perceived variance faster than perceived mean, causing agents to cut v and triggering a crash.
Main quantitative findings (with magnitudes and scope): Theoretically, the SCE is stable only for gains below a threshold; at epsilon-bar the Jacobian of the averaged system has complex eigenvalues on the unit circle (a Neimark-Sacker / discrete Hopf bifurcation), and above it a stable limit cycle exists (Theorem 1, using Kuznetsov 1998). The threshold is approximately epsilon-bar = 8.9 x 10^-4, far below the calibrated epsilon = 0.0052 (about six times larger), so empirically plausible gains imply instability. Eigenvalues at threshold: 0.512 +/- 0.859i = e^(+/-1.0333i). Calibration uses Shiller (2024) S&P 500 data, 1871-2022 annual: empirical P/D mean 28.97, sd 15.53; log P/D mean 3.25, sd 0.46; 100x log return mean 6.51, sd 16.90; dividend growth 100x(d,sigma)=(1.56, 11.104). Optimizing (beta,gamma,epsilon) the baseline matches log P/D (mean 3.15 vs 3.25, sd 0.46 vs 0.46) and returns (6.44 vs 6.51; sd 16.85 vs 16.90) with beta=0.979, gamma=3.278, epsilon=0.0052, and a low risk-free rate 100xlog R=0.87. Crashes (defined as a 30% P/D drop) occur every ~38 years in the baseline vs ~25 years in data; matching the data frequency would need a larger gain near 0.025. The closed-economy and rational-expectations versions essentially cannot produce such crashes. Drawbacks: consumption growth is too volatile (sd ~16.79 vs 1.27 in data) and return predictability is far stronger than in the data.
Layer 2: Deep Dive
What exactly drives the instability, and how is it established rather than merely simulated?
Instability comes from the feedback between beliefs (m, s) and the net asset/debt position omega: beliefs set the portfolio share, which sets prices and returns, which feed back into beliefs. Williams formalizes this by stacking current beliefs, lagged beliefs, and the state omega into a 5-dimensional first-order system X_{t+1}=G(X_t, chi_t), then studies the deterministic averaged system Xbar_{t+1}=Gbar(Xbar_t) (averaging only over the i.i.d. dividend shocks chi, NOT over omega as the small-gain limit does). Linearizing at the SCE fixed point, Theorem 1 shows all Jacobian eigenvalues lie inside the unit circle for gains below a threshold epsilon-bar, a complex pair hits the unit circle at epsilon-bar (Neimark-Sacker bifurcation), and a unique stable closed invariant curve (limit cycle) appears for epsilon just above. He verifies the nondegeneracy and stability conditions numerically.
Why does small-gain analysis mislead here, and what is the methodological contribution?
Standard learning convergence results take the gain to zero, treating state dynamics as ‘fast’ relative to beliefs and averaging over the state. Williams shows this is valid only for extremely small gains in his model because the radius of stability is tiny (epsilon-bar ~ 8.9e-4). Averaging over omega destroys the very belief-state feedback that drives cycles. His contribution to the learning literature is applying discrete-time bifurcation theory (Kuznetsov 1998) to show a Neimark-Sacker bifurcation and stable limit cycle in an economic learning model — which he states is novel — relating it to prior cautions by Cho (2018), Chien-Cho-Ravikumar (2020), and instability examples in Evans-Honkapohja (2009) and Honkapohja-McClung (2023).
What are the two crash mechanisms and which dominates?
(1) Binding leverage constraint: if v hits vbar during a boom, inflows stop, generating a negative return surprise that lowers the mean estimate and cuts v. This is rare in the calibration — it drives only the final crash in the long simulation. (2) Endogenous volatility: a rapid boom raises both the estimated mean and variance of returns; when the variance effect dominates, agents cut the risky share even without hitting the constraint. Because the economy is in the steeply sloped pricing region, a tiny cut produces a large crash. This is the dominant, novel mechanism and causes all other crashes, including those in the highlighted closeup. In one example the portfolio share peaks just above one (period 441), and a move from v=1.004 to 1.000 produces about a 48% P/D drop; the cascade bottoms near v=0.47 and P/D around 2, a decline of over 95% from peak.
What does the representative boom-bust cycle look like quantitatively?
In a >1,000-period simulation, P/D rises 30-50% within a span of years then crashes by a similar or larger amount. In the detailed cycle the P/D rises from 30 to 50 over a few periods before crashing to around 2. After a crash, volatility estimates start high and decline monotonically over roughly 50 periods; agents slowly raise v, prices rise (amplified by the omega multiplier as accumulated bonds are sold), until a rapid boom enters the fragile region and crashes again. Severe crashes of similar magnitude recur at periods 327, 442, 801, and 1067.
What is the role of stochastic shocks versus endogenous dynamics?
Conditional impulse responses (at periods 432, 438, 440 into a boom) show shocks matter most early: at t=432 a positive shock reinforces the boom while a negative shock dampens fluctuations with little belief change. By t=438 positive/negative impulses are qualitatively similar but differ in magnitude. By t=440 the endogenous dynamics dominate and shock differences are minimal — the boom continues only a couple periods before a severe crash. Shocks govern timing and magnitude, but endogenous belief changes ultimately drive the cycles.
How does the open-economy assumption matter, and what is the closed-economy comparison?
The baseline is a small open economy: international trade in bonds (fixed R) but only domestic equity trade, which permits nonzero net debt and asset flows. This debt/portfolio-adjustment channel is essential. In the closed economy (R adjusts each period to clear bonds at zero net supply, v=1), with baseline parameters the fit is much worse: P/D too high (3.70), returns lower (4.08), and far less volatile (sd P/D 0.15). Re-optimizing the closed model improves means but misses volatilities (overshoots return sd at 17.74, undershoots P/D sd at 0.36) and requires very different parameters (beta=0.903, gamma=4.736, epsilon=0.0272); crashes occur only every ~469 years (extremely rare). Intermediate cases with partial interest-rate adjustment keep the closed-economy qualitative features. The empirical justification: foreign investors held 33% of US Treasuries, 27% of corporate debt, but only 17% of US equities in 2023 (vs 46% Treasuries and 9% equities in 2006).
How does the speed of learning (gain) trade off against fit?
As the gain falls toward zero, the P/D ratio converges to its SCE value log(P/D)~3.6 and its distribution concentrates there (lower volatility); higher gains raise volatility and crash frequency but lower the mean P/D because more time is spent recovering from crashes (booms are short-lived, crashes slow to recover — an asymmetry). The calibration balances mean and volatility of P/D at epsilon=0.0052, but matching the observed crash frequency would need a larger gain near 0.025. The model can match price level/volatility OR crash frequency but struggles to match the speed of market dynamics simultaneously.
What are the main empirical drawbacks?
(1) Consumption growth is far too volatile (model sd ~16.79 vs data 1.27), inherited from using volatile empirical dividend growth as the driving process; treating stocks as levered equity claims (Abel 1999) could break the consumption-dividend link. (2) Return predictability — both autocorrelation and long-term reversal — is much stronger than in the data, where it is weak at best; additional shocks or heterogeneity would dampen it. (3) The subjective excess return is essentially uncorrelated with the P/D ratio, whereas survey expected returns are positively correlated with P/D (Greenwood-Shleifer 2014; Adam-Marcet-Beutel 2017; Barberis et al. 2018); allowing different gains for the mean and variance moves the model closer to survey evidence.
How does this differ from closely related prior work?
Versus Branch and Evans (2011), who also have agents learning about risk and return: their booms/crashes are rare ’escape’ events from equilibrium, whereas in Williams’s model they are typical outcomes driven by a fundamental instability (a stable limit cycle), not rare escapes. Versus Adam, Marcet and Nicolini (2016): Williams adds a fixed-rate risk-free asset, creating a portfolio problem and debt dynamics (omega) that are crucial for the boom-bust cycles. Versus behavioral/extrapolation and diagnostic-expectations models (Barberis et al. 2018; Bordalo-Gennaioli-Shleifer 2018; Bianchi-Ilut-Saijo 2024), Williams uses standard adaptive learning, and crucially crashes collapse valuations far below fundamentals (not mere reversion to fundamentals), with stability breeding instability as in Minsky.
What are the policy implications and their scope conditions?
A full policy analysis is outside the paper’s scope, but Williams notes a higher interest rate lowers excess stock returns and makes boom-bust cycles less frequent — yet potentially more severe (when a boom does occur, larger price/return spikes). This implies policymakers face tradeoffs more complex than simply ’leaning against the wind’ of bubbles. The scope conditions: the model has exogenous output growth, a representative agent, a constant risk-free rate, and a constant rational-expectations P/D, so all fluctuations are attributed to learning; relaxing these (e.g., for finance-real interactions) is left for future work.