Ayush Jha

Generative AI may help investors process public information, but widespread use of the same models can make their decisions more alike. This paper asks whether common AIrelated information remains underused in stock prices or has become sufficiently reflected that further reliance produces overreaction. We derive a sign restriction from a noisy rational expectations model: conditional on the price, the common signal predicts returns positively when underused and negatively when overweighted. We test this prediction around 98,461 U.S. earnings announcements from 2015 to 2025 using a point-in-time information index and stocks' exposure to retail orders routed through brokers that later introduced AI assistants. After ChatGPT, the signal predicts announcement returns positively across all exposure groups, with a significant association among the most exposed stocks. We find no exposure-related change after ChatGPT or broker launches. The evidence is consistent with continued underreaction, rather than AI-induced overreaction, in this empirical setting.

Backtests of past winners often use databases that retain firms that survive, but it is unclear whether the resulting bias reflects changes in which firms are selected or the omission of outcomes for firms that later exit. We separate these channels using point-in-time data on U.S. common stocks and annual winner portfolios formed from 1991 to 2020, with outcomes observed for up to five years. Conditioning the ranking on five-year survival changes one-third of the top twenty names, roughly the amount implied by overall attrition. The resulting survivor-conditioned portfolio nevertheless earns 6 percentage points more over one year and 37 percentage points more over five years than the corresponding full-universe portfolio. An exact decomposition attributes nearly all of this difference to the returns omitted when winning firms exit, rather than to the firms promoted into their place. The effect is concentrated among equal-weighted winners, where distress-related exits are associated with large losses.

Portfolio optimization almost universally treats the riskless rate as exogenous, typically a short Treasury yield. In incomplete markets, this is an artifact as the rate at which an investor can discharge risk depends on the hedging technology inside her own portfolio, not on a security she may not trade at scale. We formulate portfolio choice as a recursive learning operator on conditional covariance operators by selecting a minimum-risk direction, re-estimating its return law under long-memory, heavy-tailed dynamics, re-injecting it, repeat. The map admits a fixed point whose expected return defines the Adaptive Minimum-Risk Rate, an endogenous, portfolio-specific shadow benchmark. Contraction is governed by the covariance operator's condition number, destabilizing precisely when assets become near-collinear. Empirically, the recursion converges on equity and cryptocurrency universes to a fixed-point invariant to the informational regime, lying orders of magnitude below the raw minimum-variance level-separating reducible risk from the structural residual no hedge removes.

We study whether the timing of quote revisions contains information about liquidity stress that is not captured by the magnitude of price movements. We develop a subordinated-Lévy framework that separates return variation into a diffusion component and a component associated with irregular business time. The clock-dispersion parameter measures how unevenly price revisions occur over calendar time. Using S&P 500 index quotes around stress and policy events and 779 trading days of SPY National Best Bid and Offer data from 2007-2025, we find that clock dispersion is related to the dispersion of mid-quote revisions and rises sharply around FOMC statements. The response is concentrated in the statement window, while changes are not observed around CPI and employment releases. We find little evidence that clock dispersion forecasts liquidity conditions. The results support the interpretation that timing irregularity can help describe liquidity stress within an event window, while its forecasting role remains limited.

What happens when the portfolios generated by an efficient-frontier procedure become inputs to the next frontier? Standard portfolio choice takes the investable set as fixed. We instead allow the procedure to enlarge that set by adding a fitted return law for a portfolio it has selected, and repeat the optimization until the resulting frontier stabilizes. We characterize this limiting object as a fixed-point efficient frontier. At the fixed point, the first-order conditions imply a beta-pricing relation on the portfolio support and, at a self-consistent tangency point, an endogenous capital market line that does not require an externally specified riskless rate. In the Dow 30 application, the formal construction is stable across rolling windows, but a spanning placebo indicates that much of the observed risk reduction comes from the expansion of the attainable set rather than from the dependence structure of the fitted augmentation. The analysis therefore also develops an identification profile that shows where along the frontier the allocation is most sensitive to estimation error.

Standard models of misspecified learning take the agent's information structure as given. This paper allows that information structure to be endogenous. Rationally inattentive agents choose which statistics to process using the maintained model itself, so a misspecified model can direct attention away from evidence that would challenge it. In a dynamic asset market with capacity-constrained traders, the pseudo-true model is the spectral projection of the truth onto the affordable class where a persistent but transitory "hype" factor is capitalized at the permanent-growth multiple. The misspecification is not revealed within that class, which is just identified by the moments traders already track; it surfaces only in an over-identifying restriction whose evaluation costs capacity. The resulting threshold rises with the market's coordination motive, with the mass of traders in the low-monitoring state and, perhaps less intuitively, with fundamental uncertainty. Mispricing need not vanish as capacity grows, as additional capacity can reduce inattention without resolving misspecification. We apply the framework to the artificial-intelligence valuation debate.

We develop a partial integro-differential equation (PIDE) framework for option pricing under joint stochastic volatility and jump dynamics, and evaluate its empirical content using the S&P 500 index option contracts across three maturities. The framework is derived from the infinitesimal generator of an affine Lévy-type process and implemented via finite-difference discretization with FFT-based treatment of the nonlocal jump operator. Calibration via GMM reveals that stochastic volatility accounts for the dominant share of pricing improvement, where relative to Black-Scholes, the Heston specification reduces implied-volatility RMSE by 39%. Jump augmentation via either Merton or CGMY specifications yields marginal improvements concentrated at short maturities and in the deep out-of-the-money region. The calibrated CGMY activity index supports a compound-Poisson structure, consistent with high-frequency evidence on S&P 500 index returns.

Professional forecasters report full predictive distributions, yet most research on expectation formation focuses on point forecasts. We use density forecasts from the Survey of Professional Forecasters to study how subjective beliefs differ from benchmark predictive distributions constructed in real time from statistical forecasting models. Comparing survey densities with these benchmark distributions , we construct empirical weighting functions that characterize how forecasters allocate probability across future outcomes. The results reveal substantial heterogeneity across macroeconomic variables. Forecasts for real GDP growth and core inflation exhibit pronounced S-shaped weighting functions, indicating that forecasters assign relatively less probability to extreme outcomes and more probability to outcomes near the center of the distribution. In contrast, forecasts for GDP deflator inflation and the unemployment rate remain close to the linear benchmark. The results are robust across alternative benchmark models, sample periods, and forecast horizons. Overall, the evidence suggests that professional forecasters systematically compress tail risks for some macroeconomic variables but not for others, highlighting an important dimension of expectation formation that is not captured by point forecasts alone.

This paper examines how social media functions as an investment-information channel and how platform-specific usage relates to speculative retail trading. Using nationally representative household data from the 2024 U.S. National Financial Capability Study, we link investor-reported use of individual social media platforms for investment information to selfreported meme-stock participation. Estimating platform-by-platform logit models and reporting average marginal effects, we document a pronounced platform gradient in speculative trading. Engagement with community-driven, discussion-oriented platforms - most notably Reddit, Discord, and Telegram - is consistently associated with a higher probability of meme-stock purchasing. In contrast, broadcast-style or professionally oriented platforms such as YouTube and LinkedIn exhibit economically small and statistically insignificant effects, while several mainstream platforms display weaker and imprecise associations. Complementary analyses using doubly robust estimation, heteroskedasticity-based instrumental variables following Lewbel (2012), and Oster-style sensitivity bounds indicate that these platform effects are consistent with causal transmission rather than purely selective participation. Reliance on financial influencers does not predict speculative participation once platform exposure is accounted for. The findings highlight the central role of peer interaction and collective narratives - rather than one-to-many persuasion - in shaping speculative retail behavior, with direct implications for investor protection and platform regulation in digital financial markets.

Rational expectations are well-supported in financial markets. Evidence that traders are also rationally inattentive, however, runs through a single channel: the business cycle. Whether inattention binds independently of cyclical variation-as a primitive cognitive friction mapping objective events into subjective beliefs-remains open. We answer using NFL prediction markets: binary contracts traded on Polymarket against a cash-like safe asset, benchmarked against ESPN's continuous win probability, in an environment that trends but does not cycle. Using 2025-26 season data and a dynamic program nesting FIRE and cognitive discounting, we reject FIRE three ways: U-shaped pre-game volatility drifting with competitiveness, a persistent state-dependent wedge unaltered by concurrent market load, and precisely zero cross-game spillovers. These collapse to a cognitive discount of 80%, validated by a reinforcement-learning agent reproducing the inaction and panic-liquidation patterns. The result isolates acyclical, purely-cognitive rational inattention, locating the friction at the state-to-probability mapping rather than the attention-across-assets margin.