Yinan Su

We model predictive scale-specific cycles. By employing suitable matrix representations, we express the forecast errors of covariance-stationary multivariate time series in terms of conditionally orthonormal scale-specific basis. The representations yield conditionally orthogonal decompositions of these forecast errors. They also provide decompositions of their variances and betas in terms of scale-specific variances and betas capturing predictive variability and co-variability over cycles of alternative lengths without spillovers across cycles. Making use of the proposed representations within the classical family of time-varying conditional volatility models, we document the role of time-varying volatility forecasts in generating orthogonal predictive scale-specific cycles in returns. We conclude by providing suggestive evidence that the conditional variances of the predictive return cycles may (i) be priced over short-to-medium horizons and (ii) offer economically-relevant trading signals over these same horizons.

We estimate a narrative factor pricing model from news text of The Wall Street Journal. Our empirical method integrates topic modeling (LDA), latent factor analysis (IPCA), and variable selection (group lasso). Narrative factors achieve higher out-of-sample Sharpe ratios and smaller pricing errors than standard characteristic-based factor models and predict future investment opportunities in a manner consistent with the ICAPM. We derive an interpretation of the estimated risk factors from narratives in the underlying article text.

Whether, how, and which firm characteristics determine the cross-sectional variation of expected stock returns? We develop a latent factor model with time-varying loadings (Instrumented Principal Components Analysis or IPCA) which allows observable characteristics as instruments for the unobservable dynamic loadings. IPCA infers that characteristics affect expected return by driving the exposure to latent risk factors, and rules out characteristics-associated anomaly (compensation without risk). Four IPCA factors explain the cross-section of average returns significantly more accurately than existing factor models. Furthermore, among a large collection of characteristics explored in the literature, only eight are statistically significant in the IPCA specification and are responsible for nearly 100% of the model's accuracy.

We propose a new model of expected stock returns that incorporates quantity information from market trading activities into the factor pricing framework. We posit that the expected return of a stock is determined by not only its factor risk exposures (β) but also the factor's quantity fluctuations (q) induced by trading flows, and hence term the model beta times quantity (BTQ). The rationale is that sophisticated investors should demand a higher factor premium when they have absorbed noise trading flows of stocks with high loadings to that factor. The BTQ model provides a compelling risk-based explanation for stock returns, which is otherwise obscured without considering the quantity information. The cross-sectional risk-return association, which is nearly flat unconditionally, strongly depends on the quantity variable. The structured BTQ model reliably predicts monthly stock returns out of sample, and addresses the factor zoo problem by selecting a small number of factors.

We propose a new approach of latent factor analysis that, in addition to the main panel of interest, introduces other relevant data that serve as instruments for dynamic factor loadings. The method, called IPCA, provides a parsimonious means of incorporating vast conditioning information into factor model estimates. This improves the efficiency of estimates for the latent factors and their loadings, and helps to ascertain the economic relationships among factors and individuals via the observable instruments. The estimation is fast to calculate and accommodates unbalanced panels. We show consistency and asymptotic normality under general panel data generating processes. We demonstrate the advantages of IPCA in simulated data and in applications to equity asset pricing and international macroeconomics.

Portfolio optimization chiefly focuses on risk and return prediction, yet implementation costs also play a critical role. Predicting trading costs is challenging, however, since costs depend endogenously on trade size and trader identity, thus impeding a generic solution. We focus on a key, yet general, component of trading costs that abstracts from these challenges—trading volume. Trading volume is highly predictable, especially with machine learning. We model the economic benefits of predicting volume through a portfolio framework that trades off portfolio tracking error versus net-of-cost performance—in essence translating volume prediction into net-of-cost portfolio alpha. We find the benefits of predicting volume to be substantial, and potentially as large as those from return prediction.

We estimate a cross-sectional asset pricing model with structural breaks and regime-dependent momentum trends in factor premia. The method incorporates total-variation regularization (TV) in stock return prediction to estimate a time-varying slope factor model. It produces piecewise linear trend lines of varying frequencies, capturing either long-term or fine-grained factor momentum trends depending on the TV penalty, which is tuned in a machine learning framework. Large-sample theoretical properties of the estimators include consistency and asymptotic normality. OOS trend-following factor-timing strategies yield superior performance compared to buy-and-hold or traditional momentum strategies with fixed windows. Time-varying factor selection and premia estimates yield improved SDF construction compared to the static counterparts.

Compute (computing power) is a scarce, capital-intensive input at the center of the AI economy. Compute capital expenditure and service flow already exceed 1% of U.S. GDP and are growing rapidly. The price of compute reflects uncertainty over AI adoption. The announced launch of compute futures turns this uncertainty into a tradable risk, raising questions on the pricing of a new asset class. We provide an early asset-pricing framework for compute. We begin by discussing the underlying compute rental market and its indexation. We then turn to pricing: 1) direct no-arbitrage links between futures prices and current spot prices fail due to the non-storable nature of compute, 2) synthetic futures prices from existing term rental contracts are likely upper bounds on true futures prices and, 3) upon financialization, futures prices will be investors' expectations of spot prices at expiration net of a risk premium. Using synthetic futures as stand-ins before the compute futures market launches, we construct the first compute futures return panel sorted by GPU generation and maturity. Our preliminary evidence is consistent with a positive compute risk premium, suggesting hedging pressure on the part of compute providers.

Interbank lending is beneficial but subject to coordination failure. With interbank wholesale funding, banks' balance sheets become inflated, which give the retail depositors a sense of safety to allow the bank to have more illiquid investments. In interbank runs, banks run on banks as they mutually reinforce each other to withdraw interbank lending. Banks' individually precautionary liquidity hoarding strategies are connected by the pairwise lending relationships. Mean-field analysis extracts the systemic behavior from the network of strategic interactions. I show such dispersed and indirectly linked interactions also lead to discontinuous and system-wide liquidity crunches as if the interactions are centralized. Local insolvency shocks trigger the interbank run if the network is unraveled beyond a critical point. The model is applied to identify the optimal capital injection targets of government bailouts and study the systemic effects of the proposed regulations on restraining the highly connected banks.

This paper studies the general equilibrium effects of industry-specific productivity shock in an economy in which sectors are connected via input-output linkages. My central finding is productivity shocks do not only travel downstream as is standard in the literature, but also trigger demand change at the final consumption industries, which propagates upstream. I label this novel mechanism "reflection channel". Differences of the elasticity of substitution of consumption and production for the final consumption industries drive the demand change. Empirically, the magnitude of the reflection channel is around three times greater than the previously studied downstream channel. When a positive productivity shock reaches a final consumption industry, consumers substitute towards it much more than producers substitute away, increasing the demand of its upstream industries, and vice versa.