The tradeable claim is the skew reversal. Below peak volatility, calls are rich; above it, puts are rich. Yet the paper supports that mechanism only with Monte Carlo at a single maturity. Nothing here establishes whether a filtered growth belief improves a listed index surface.
The central theoretical result is negative. Within the authors' candidate class, equilibrium eliminates any extra absolutely continuous valuation state. The scope is tight: the added factor must be the unique global positive solution of an ODE driven by (beliefs, factor), and every initial belief must use the same Borel maps.
Testing the option claim is our question, not theirs. We have no results yet. Any test we could run substitutes SPY for the continuously observed aggregate dividend in the paper and replaces the exact equilibrium with an estimated discrete daily hidden-Markov model. Such a test would cover an approximation of the mechanism. It could never test the paper's pricing levels.
Filtering growth under Epstein-Zin preferences
Aggregate dividends are geometric, with constant volatility sigma. Their drift switches among n hidden states according to a Markov chain. Investors observe dividends alone and infer the state through the Wonham equation, the filtering recursion for a chain seen through a diffusion. The posterior belief vector becomes a state variable alongside the dividend level.
Preferences are infinite-horizon Epstein-Zin, represented by a BSDE, with theta = (1 - gamma)/(1 - 1/psi). A representative agent clears the market in equilibrium.
Pagès, Possamaï and Rodriguez Polo allow more price dependence than a belief-based price-dividend ratio. Their candidate price is S = D H phi(P), where phi is a smooth positive function of beliefs. The positive, absolutely continuous process H follows its own ODE driven by (P, H), giving it scope to retain the history of beliefs.
Their Theorem 3.1 removes that extra state. Given a classical C^2 state representation of the value function and a positivity condition on the coefficient zeta, one constant h0 > 0 determines every initial belief: h(p) = h0, while H_t = h0 for all t. Consequently, Q = h0 phi(P), and price is Markovian in dividends and beliefs. As the authors put it, "Markovianity is an equilibrium conclusion, not a restriction imposed on candidate prices."
The abstract states the boundary clearly, promising the result only "within the class C defined below and under the regularity, admissibility, and positivity conditions of our main theorem". The live issue is therefore the reach of that class.
The proof mechanism is instructive. The Hamilton-Jacobi-Bellman equation and local optimality condition still permit non-Markovian candidates. Set h(p) = a above a threshold, and a classical solution appears with pointwise optimal controls pi = 1 and c = x/(u phi(p)); both clear pointwise at market states.
Admissibility kills the candidate.
The ratio also lacks valuation stability. Since Q grows at least as fast as exp(zeta_min t), sup_t E[Q_t] is infinite. Local optimality alone cannot identify the equilibrium. In the general case, positivity of zeta remains an assumption. For two states with strictly positive intensities, the paper proves it, clearing that obstacle from the two-state results.
The boundary-value problem with two states
With n = 2, the pricing equation reduces to a scalar ODE. It degenerates at both endpoints because the belief diffusion chi(p) = p(1-p)Δg/sigma vanishes there. Suppose c(p) > 0 on [0,1] and both transition intensities are strictly positive. For every theta > 0, the authors prove the existence of exactly one positive classical solution satisfying theta/c+ <= phi(p) <= theta/c-. The solution is smooth across the closed interval and real analytic in its interior.
Existence uses a positive compact resolvent for the linear component, followed by a Schauder fixed point on a truncation. The a priori bounds make that truncation inactive. A ratio maximum principle supplies uniqueness.
Preferences determine the shape. Monotonicity follows the sign of eta = (1 - gamma)(g1 - g2), while theta governs curvature. The profile is convex for 0 < theta < 1 and concave for theta > 1, provided a coefficient m_theta stays positive. That coefficient depends on the unknown solution through 1/phi.
At theta = 1, the ratio becomes exactly affine: phi(p) = a + bp, where b = eta/(c(0)c(1) + c(0)lambda12 + c(1)lambda21). No additional condition is needed. The authors also derive sufficient conditions stated entirely in parameters, then show in their own Figure 2 that those conditions are conservative. Its parameters are g1 = 0.21, gamma = 0.75, sigma = 0.49, lambda12 = 0.052, lambda21 = 0.059, delta = 0.049, with theta in {5/6, 2}. Both primitive inequalities fail. Even so, m_theta remains positive and the convex and concave shapes survive.
The equilibrium interpretation extends only to 0 < theta <= 1. For theta > 1, the results describe the ODE's positive solution, pending a utility-selection rule that the paper does not construct. The authors explicitly leave theta < 0 unanalyzed and observe that this region "includes the often-used parameter combination gamma > 1 and psi > 1".
Why the skew changes sign
Stock volatility is sigma_S(p) = sigma + l'(p)chi(p), where l = log phi. A single innovation changes cash flows and beliefs together, so rising phi adds learning-generated volatility. At short maturities, the leading conditional risk-neutral skewness is 3 chi(p) sigma_S'(p)/sigma_S(p) times sqrt(tau). Its sign follows the direction in which volatility responds to a belief revision. A unimodal sigma_S therefore makes the skew reverse at the volatility peak.
The option calibration sets g1 = 0.05, g2 = -0.05, sigma = 0.04, delta = 0.06, lambda12 = lambda21 = 0.05, gamma = 0.8, theta = 0.25, psi = 5 and tau = 0.10. Volatility peaks at p* = 0.486. At the degenerate beliefs, stock volatility equals dividend volatility exactly, at 4%. Inside the interval it reaches about 35.6%.
Nearly nine times higher.
Essentially all of that volatility comes from learning, a consequence of pairing a 10-point growth spread with 4% dividend volatility. Over tau = 0.10, simulated risk-neutral log returns have sample skewness +0.98 at p0 = 0.20, where sigma_S'(p0) = +0.76. At p0 = 0.80, sample skewness is -0.99 and sigma_S'(p0) = -0.75. Their Figure 5 shows a risk reversal moving from about -10 to +10 volatility points across beliefs. Those readings use tau = 0.10 and |log(K/F)| = 0.07, with relative wing premia running between about -100% and +150%.
Below the volatility-maximising belief, the model makes out-of-the-money calls rich. Above it, downside puts become rich. This is the paper's one implication I would trade on.
The sign changes at 0.486 in the calibration. With symmetric intensities lambda12 = lambda21 = 0.05, the model spends roughly half its time on either side. Index surfaces, however, have remained persistently put-skewed since 1987. The authors address this directly, writing that the sign-changing skew "should therefore be read as a state-dependent mechanism". Standard index put skew belongs to states in which downside news raises future volatility. Call-rich surfaces emerge during recoveries, when upside news moves beliefs toward the high-volatility region.
As an interpretation of the states, the account hangs together. It does not generate a belief process that remains above p* as often as the data would require. The paper's most distinctive prediction is also its easiest one to falsify.
Calibration all the way down
Every numerical result in the paper comes from calibration or simulation. The two-state baseline sets g1 = 0.25, gamma = 0.8, sigma = 1.25, lambda = 0.04, delta = 0.025 and theta = 1. Across twelve paths over 120 time units, the ratio remains inside [7.2, 9.2], with a long-run conditional mean of 8.2. If one time unit represents a year, sigma = 1.25 means 125% dividend volatility.
For the option results, the authors use 512-step Euler paths, antithetic increments and Black control variates. Smiles and densities use 1,048,576 paths; risk reversals use 262,144. We did not find standard errors or convergence diagnostics for those prices. The paper describes the parameter choice plainly: "This is a stylised illustration of the mechanism, not an empirical calibration."
Three further restrictions stand between the skew plot and an empirical prediction. First, equilibrium leaves the short rate incompletely determined, requiring only R <= r(p). Every option calculation chooses the marginal upper bound, which spans 1.17% to 6.92% in the calibration. Second, the option section assumes eta > 0, equivalently gamma < 1, and uses gamma = 0.8. Third, the skewness formula is an expansion for return skewness. The paper says it "does not by itself establish the slope of implied volatility". Implied-volatility slope appears only in Monte Carlo at tau = 0.10.
Can filtered beliefs forecast SPY skew?
We judged the mechanism buildable, and a strategy is under evaluation. We have no results yet. The latent growth states, preference parameters and marginal short rate cannot be observed, so we estimate a discrete daily hidden-Markov model and calibrate the remaining quantities numerically. ETF distributions provide only a coarse substitute for a continuously observed dividend process. Options observations arrive at end of day, while hedging and trading can occur only once a day.
The filtered belief must survive a direct horse race. It needs to predict changes in the SPY risk reversal after controlling for lagged implied skew and the level of implied volatility. It must also outperform a plain two-state regime-switching benchmark. Without that evidence, the reversal at p* = 0.486 remains a feature of this calibration rather than a signal.
A time-consistent belief series would change my mind if its slope term chi(p)sigma_S'(p) forecast the next-day sign of the SPY 25-delta risk reversal, with incremental significance over the lagged risk reversal. The equilibrium theorem stands as a real contribution on its own. The option hypothesis still has to face a price series.