Abstract We develop a network-based vector autoregressive approach to uncover the interactions among financial assets by integrating multiple realized measures. Under a restricted parameter structure characterized by a single integrated multimeasure network, our approach captures cross-sectional and time dependencies embedded in a large panel of assets. We propose a block coordinate descent procedure for the least square estimation, investigate its theoretical properties and assess estimation consistency through simulations. Using data on U.S. stocks, we identify a large array of interdependencies with a limited computational effort. We provide a new ranking for the systemically important financial institutions and carry out an impulse-response analysis to quantify the effects of adverse shocks on the financial system. Exploiting information through the integration of multiple realized measures, our method leads to significantly more accurate out-of-sample forecasts, compared to single measure models and competing Principal Component Analysis (PCA)-based factor model specifications.
Abstract In this study, we introduce smooth time-varying coefficients to the three-pass regression filter (3PRF) forecasting method. We employ a rolling window selection procedure to estimate the parameters and to generate forecasts. The rolling window selection uses only the most recent observations, which resolves the trade-off between forecast bias and variance. We establish the optimal rate for selecting the most recent observations. Monte Carlo simulations demonstrate that, in general, the rolling window selection method for 3PRF forecasting with time-varying coefficients produces relatively smaller forecasting mean squared errors compared to the original 3PRF forecasting method or the method that uses the full sample. Empirical applications of our proposed method are considered to forecast eight macroeconomic variables and market returns, which highlight the necessity of using our approach.
When factors for asset pricing are constructed from a set of sorted portfolios, statistical tests of factor model alphas can suffer from low power. The low power arises because alpha can be decomposed into a sum of two parts, where each part is equal to zero under the null of a correct factor model. If the values of the two parts under the alternative hypothesis are nonzero, differ in sign, and have approximately equal magnitudes, they offset which results in a high probability of not rejecting a false model. Thus, insignificant alpha tests can be misleading when factors are constructed from sorted portfolios. This paper uses the parts of alpha to derive new test statistics that have high power when the traditional alpha test lacks power.
Abstract We examine the channels and efficacy of monetary policy at the zero lower bound (ZLB) through the lens of various shadow rate models. Our key methodological contribution is to extend the discretization filter to incorporate missing observations. This allows us to estimate shadow rate models that both incorporate survey forecasts and allow for departures from rational expectations. The models disagree about the level of the shadow rate and the duration of the ZLB in real time, but they are remarkably consistent in terms of their implications for the effects and channels of policy across a number of structural and reduced-form exercises. Particularly, they attribute most of the effects of major Federal Reserve policy announcements to changes in term premia, and imply that large scale asset purchases affected term premia both by changes in average duration and changes in local supply of substitute assets.
We analyze jump risks in financial asset prices modeled by Ito semimartingales with an exponential Hawkes process as the jump counter. First, using little information, we estimate the probability that an observed jump cluster is not yet exhausted. Second, we make explicit the conditional density of consecutive jump durations and prove that durations stochastically increase. Third, we provide bounds for jump probabilities in consecutive time intervals. Application to 5-minute U.S. returns shows that cluster depletion probabilities strongly correlate with the expected yearly jump count, and improve jump forecasts.
We derive an admissible parameter space for vector multiplicative error models (vMEMs), explicitly formulating it in terms of the model's matrix parameters through a set of matrix inequalities. Another key contribution is the adoption of constrained maximum likelihood estimation for the multivariate process, which ensures compliance with these matrix inequalities and addresses the limitations of unconstrained approaches used in previous studies. To demonstrate the effectiveness of the proposed method, we apply it to four empirical cases in financial volatility modeling, emphasizing its practical relevance.
Model calibration is a challenging yet fundamental task in financial engineering. Using sequential Monte Carlo methods, we reformulate the nonconvex optimization problem as a Bayesian estimation task. This allows to compute any statistic of the estimated parameters, mitigating the strong dependence on starting points and avoiding the troublesome local minima, that plague standard calibration methods. To accelerate computation, we incorporate Markov chain Monte Carlo methods with delayed acceptance and a neural network-based option pricing approach. When applied to S&P 500 index options, our Bayesian algorithms significantly outperform the standard approach in terms of runtime, accuracy, and statistical fit.
We investigate the benefits of including hedge funds into a portfolio of stocks, bonds, and commodities. We use a multivariate canonical vine copula regime-switching model which allows for non-linearity, asymmetry, and time variation in hedge fund returns. We find that the willingness to pay to access hedge funds is about 4 cents per dollar, and it increases with risk aversion; the weights in hedge funds show an inverse U-shape with risk aversion; hedge funds tend to replace stocks (bonds) for risk-averse (risk-tolerant) investors; investing in hedge funds increases historical returns only until 2008, but reduces volatility even after.
We propose a novel stochastic volatility model based on the autoregressive gamma process that accommodates a structure-preserving change to the risk-neutral measure while relying on a non-Gaussian distribution for the return innovations. The model employs the Meixner (MXN) distribution, which enriches the return dynamics with conditional stochastic skewness and kurtosis. We propose a fast and accurate estimation method by combining the approximate maximum likelihood method of David S. Bates with a numerical integration technique suitable for highly oscillatory functions. We derive a closed-form discrete-time option pricing formula. The MXN model performs particularly well, compared to benchmarks within its class and of the generalized autoregressive conditional heteroskedasticity family, when calibrated directly to option data and when applied to option data with a high level of implied volatility, such as Bitcoin.
We introduce a new type of shrinkage estimator that is not based on asymptotic optimality, but instead learns a state-dependent shrinkage policy via supervised learning in a contextual bandit setup. The proposed estimator applies to both linear and nonlinear shrinkage and shows improved performance compared to classical shrinkage estimators. Our results demonstrate that our estimator identifies a downward bias in classical shrinkage intensity estimates derived under the i.i.d. assumption and automatically corrects for it in response to prevailing market conditions. Additionally, our data-driven approach enables more efficient implementation of risk-optimized portfolios and is well-suited for real-world investment applications, including portfolios with practical optimization constraints.
This article addresses practical issues associated with the use of the local estimator in forecasting models that are affected by parameter instability. We propose an approach to select the bandwidth parameter in the context of out-of-sample forecasting. Derived by minimizing the conditional expected end-of-sample loss, the selection procedure is shown to be asymptotically optimal. We also discuss the implications of the choice of kernel functions. The theoretical properties are examined through an extensive Monte Carlo study. Two empirical applications on forecasting excess bond returns and the yield curve demonstrate the superior forecasting performance of the local estimator with the proposed optimal bandwidth selection.
We contribute to the growing literature on high-frequency spot volatility estimation by deriving a new integral representation for the recently introduced asymptotic minimum risk equivariant (AMRE) candlestick-based class of estimators. Our new theoretical representation enables the practical numerical computation of the hitherto impractical to compute optimal estimators based on multiple adjacent candlesticks. We also propose a new exact sampling scheme for high-frequency candlestick data, which facilitates straightforward calculation of the asymptotic risk and confidence intervals for the estimators. The resulting critical values for the highest-density intervals highlight the substantial efficiency gains from incorporating more than one candlestick in the estimation process. We showcase the practical value of the new techniques in elucidating the behavior of financial market volatility around the time of important news announcements.
We develop deep neural networks with economically motivated restrictions that are designed to overcome the main challenges of factor timing. Our critical innovations include integrating multitask (MT) learning to capture the common structure across factors, with long short-term memory neural networks to extract financial and macroeconomic states. This dynamic MT neural network outperforms all benchmarks in terms of predictive accuracy and economic gains. We pinpoint unemployment, along with variations on leverage, profitability, and money as key predictors, and highlight the importance of capturing their nonlinear interactions. Improved factor timing through neural networks with economic restrictions facilitates more reliable investigation into the economic mechanisms driving factor risk premia, and underscores the value of deep learning for factor investing.
We show that the mixed causal-noncausal vector autoregressive (VAR) processes satisfy the Markov property in both calendar and reverse time. Based on that property, we introduce closed-form formulas of forward and backward predictive densities for point and interval forecasting and backcasting out-of-sample. The backcasting formula is used for adjusting the forecast interval to obtain a desired coverage level when the tail quantiles are difficult to estimate. A confidence set for the prediction interval is introduced for assessing the uncertainty due to estimation. We also define new nonlinear past-dependent innovations of mixed causal-noncausal VAR models for impulse response function analysis. Our approach is illustrated by simulations and an application to the joint analysis of oil prices and real gross domestic product (GDP) growth rates.
Tail expectations have recently attracted much attention in economics for their ability to capture risk. We develop a semiparametric estimator for the joint estimation of (nonlinear) models of tail expectations with some tail quantile as the left or right threshold, and interquantile expectations, partial expectations between two thresholding quantiles. The joint estimator of these quantities can be used to test for heterogeneity in the conditional distribution, with special attention to distinct tail behavior. We derive efficient weights and asymptotic properties of the estimator for time-series data. The estimator does not require the specification of the conditional distribution, and its computation relies on standard techniques. In an empirical application in finance, we test for a disproportionate contribution of tail events to the average abnormal return of portfolio strategies.
We propose the stepwise Cauchy combination test (StepC), a new procedure for multiple testing with dependent test statistics and sparse signals. Unlike the global version, StepC pinpoints which p-values drive rejections, while maintaining strong familywise error control. It is less conservative under dependence and more powerful than conventional multiple testing corrections. In simulations and in applications to drift burst detection and testing for nonzero alphas, StepC consistently boosts power and yields more meaningful rejections, making it a practical alternative for large-scale financial datasets.
We introduce a consistent method for pricing panels of options through time with a single simulation. The approach is particularly useful for pricing options with early exercise features and allows calibrating flexible option pricing models to large panels of such options. We showcase this by calibrating option pricing models with time-varying volatility and asymmetric features to a sample of more than 25,000 American-style options. Our method leverages the homogeneity of option prices, relies only on polynomial approximations using simulated paths, and could be applied to other problems that require estimating optimal stopping times using dynamic programming.
Extreme value regression offers a convenient framework to assess the effect of market variables on hedge funds tail risks, proxied by the tail index of the cross-section of hedge funds returns. However, its major limitation lies in the need to select a threshold below which data are discarded, leading to significant estimation inefficiencies. In this article, our main contribution consists in introducing a method to estimate simultaneously the tail index and the threshold parameter from the entire sample at hand, improving estimation efficiency. To do so, we extend the tail regression model to non-tail observations with an auxiliary splicing density, enabling the threshold to be internally determined without truncating the data. We then apply an artificial censoring mechanism to decrease specification issues at the estimation stage. Empirically, we investigate the determinants of hedge funds tail risks over time, and find a significant link with funding liquidity indicators. We also find that our tail risk measure has a significant predictive ability for the returns of around 25% of the funds. In addition, sorting funds along a tail risk sensitivity measure, we are able to discriminate between high- and low-alpha funds under some asset pricing models.
We provide a novel perspective on changes in global long-term interest rates using a dynamic term structure model of Canadian nominal and real yields with adjustments for term, liquidity, and inflation risk premiums. For the period from 1996 to 2021, we find that the steady-state or equilibrium short-term real interest rate fell by more than 4 percentage points, long-term inflation expectations edged down modestly, and real bond and inflation risk premiums varied with little longer-run trend. In contrast, our yield decomposition attributes the post-pandemic rise in interest rates largely to a sharp reversal and increase in the equilibrium real rate.
Expectiles have recently received considerable attention due to their coherence as a measure of tail risk. Estimating conditional expectiles (CExps), particularly at both intermediate and extreme levels, is crucial in quantitative risk management. This article proposes an ARMA-GARCH model that accommodates fewer finite moments and assumes innovations follow a Pareto-type tail distribution. We apply the two-step self-weighted procedure of He et al. (2022) to forecast extreme CExps. Employing extreme value theory, we estimate the extremal CExp and develop a unified asymptotic theory for the CExp estimator, which incorporates both intermediate and extreme scenarios. Our Monte Carlo simulations demonstrate that the proposed approach significantly improves coverage probabilities compared to other competing methods across various contexts, especially in extreme scenarios. Finally, an empirical application to the daily negative log-returns of fourteen financial asset indices shows that our method consistently outperforms forecasts from Hoga (2022) and the Peaks Over Threshold approach during both normal and crisis periods.
