Working Papers
Hazardous Analysts: Reputation Management Incentives and the Duration of Recommendations with Dan Bernhardt
Prior literature shows that analysts upwardly bias recommendations to generate trades, preserve investment banking relationships, or curry favor with management. It has been posited that reputation concerns of analysts would mitigate the impact of these conflicts of interest. This paper investigates this premise, using a semiparametric competing risks hazard model that accommodates both unobserved heterogeneity and the time-varying effects of covariates on analysts' recommendations to uncover the determinants of sell-side analyst recommendation revisions.
Using a comprehensive sample of analyst recommendations from 1993 to 2000, we document how incentives to manage appearances for their customer audiences appear to drive analyst recommendation revisions. In particular, we find that analysts tend to keep past good "accurate'' recommendations for too long, and to drop past bad "inaccurate'' recommendations too quickly. That is, incentives to maintain a 'good' reputation---a good-looking recommendation list in front of customers---appear to aggravate rather than alleviate conflicts of interest.
We find that recommendation revisions primarily reflect retrospective considerations, i.e., the performance of analysts' outstanding recommendations (i.e., the consistency of the recommendation with the return since the recommendation was made), rather than prospective considerations about future returns. The impact of performance on recommendations is significantly time varying: initially both very poor and very good performance raise the hazards of downward revisions, but as the duration of a recommendation increases, cumulative returns only have monotonic negative effects on the hazards of recommendation downgrades. These findings highlight the inappropriateness of both single-period classification models such as probits, and of proportional hazard models that only account for duration dependence via a time-varying baseline hazard. More generally, we exhaustively characterize the impacts of covariates on recommendation changes, for example, showing that the information content of earnings announcements is fully incorporated into recommendations within one week, and that the likelihood of revisions falls sharply with experience in an analyst's first three years (the odds of a revision in an analyst's first year are roughly three times that of an analyst with three years experience), but then plateaus.
We then ask: how does a recommendation's past performance influence the sensitivity of recommendations to future returns? Overall, analysts appear prescient: downgrades (upgrades) are more likely when future returns are bad (good). However, past poor performance of a recommendation, e.g., past very negative returns following a strong buy, dramatically reduce the sensitivity of recommendations to future returns. In particular, if since a strong buy or buy recommendation was issued, returns were highly negative, then downgrades are very likely, independently of whether future returns are low or high. In contrast when past returns were higher, the likelihood of a recommendation revision becomes much more sensitive to future returns, implying that an analyst's recommendation decision (the maintainence of a strong buy versus a downgrade) contains more information about future returns when past returns have been higher. The sensitivity of upgrades from buy and hold recommendations to future returns is more subtle, as it is time varying. Very negative past returns especially for longer-held recommendations mean that the probability of a revision is roughly orthogonal to future returns; while qualitatively, upgrade probabilities are most sensitive to future returns when past returns have not been too high.
Contagious Default
Driven by the corporate default data for U.S. public firms, I develop a method to estimate corporate default intensities that can incorporate both contagious defaults---defaults correlated through the network of contractual or financial relationships---and informatively-censored events such as mergers and acquisitions---an option to avoid bankruptcy for distressed firms. In my semiparametric hazard model, default correlations among firms and through time are modeled by two-dimensional penalized tensor-product splines. Dropouts caused by mergers and acquisitions, are treated as competing risks that are correlated with corporate defaults through a latent random variable on which they have different loadings. My approach can also be interpreted based on a mutually exciting multivariate point process model, where the direct impact of one firm's default on other firms is quantified by the financial distance and where defaults from external sources are modeled in the dependence of observable covariates and unobservable frailties. My results show that the proposal model is able to estimate the correlated corporate defaults in a more realistic way and predict the firms' credit risks with higher confidence level than the dynamic frailty model of Duffie et al. (2009), which implicitly accounts for excessive correlation caused by contagion and informatively-censored events via a common unobservable macroeconomics risk factor.
Is Cash Flow Really King? Earnings Quality and Stock Returns (September 2008)
Past research indicates that net operating cash flows are more persistent than accruals (the difference between accounting earnings and operating cash flows); firms with high operating accruals earn lower average returns than firms with low operating accruals. More recent research finds that once cash flows are accounted for, the negative association between accruals and future abnormal returns vanishes. In this paper, I analyze the complex relation between accruals, operating cash flow and subsequent returns using semiparametric and nonparametric regression approaches. I uncover a highly nonlinear and non-monotonic relationship between accruals and operating cash flows, indicating that one cannot just incorporate the impact of cash flows on abnormal future returns and ignore accruals. In particular, in the long run, accruals contain more information about future abnormal returns, while in the short run, cash flow "is king'', and has more predictive power about future abnormal returns. My findings suggest that the assumptions made by previous researchers are not valid, and that their analysis cannot appropriately capture the underlying relations, and can give rise to misleading inference. I also attempt to investigate the underlying mechanism of the accruals anomaly by determining whether and how sophisticated information intermediaries such as financial analysts use the information contained in accruals and cash flows in their earnings forecasting process. I find that that analysts are over-optimistic about firms reporting high accruals and do not correctly assess the quality of earnings. These results favor a behavioral explanation for accrual anomaly.
WORK IN PROGRESS:
Liquidity Dynamics and Stock Returns: a Bayesian Nonparametric Approach
This paper explores the joint dynamics of liquidity risks, default risks, volatility, and stock returns using a gaussian process dynamical model. Employing the flexibility of this statistical model that accommodates nonlinear and nonstationary time series, I comprehensively investigate the co-movements and interrelations of liquidity, default, volatility and returns. The results demonstrate that the intertemporal associations between those variables are not stationary, and that they vary predictably over time and with different market conditions.
After Filing Chapter 11
I investigate the hazards of mergers, emergences and liquidations after firms file Chapter 11 bankruptcy using a quantile survival analysis model. Applying this semiparametric duration model, I find considerable differences in the determinants of mergers, emergences and liquidations. Such differences reflect the primary cause for the bankruptcy---whether it is triggered by operational risks, funding liquidity risks, or market liquidity risks, etc.. Different investment values could be assigned according to these characteristics of the firms at the time when they file Chapter 11 bankruptcy. These findings shed light on the nature of optimal trading strategies of distressed stocks.
Multi-Period Corporate Default Prediction Using Joint Model of Accelerated Failure Time and Longitudinal Data
In this paper, in order to take into account the time-varying default hazard and correlation between the bankruptcies, I explore the joint modeling approach under accelerate failure time assumption by maximizing the joint likelihood function of the bankruptcy time and longitudinal data with random effects (frailty). This model produces more accurate out-of-sample forecasts than alternative models by incorporating the dynamics of both observed and unobserved (frailty) covariates. A Monte Carlo EM algorithm is used to estimate all unknown parameters.
Prior literature shows that analysts upwardly bias recommendations to generate trades, preserve investment banking relationships, or curry favor with management. It has been posited that reputation concerns of analysts would mitigate the impact of these conflicts of interest. This paper investigates this premise, using a semiparametric competing risks hazard model that accommodates both unobserved heterogeneity and the time-varying effects of covariates on analysts' recommendations to uncover the determinants of sell-side analyst recommendation revisions.
Using a comprehensive sample of analyst recommendations from 1993 to 2000, we document how incentives to manage appearances for their customer audiences appear to drive analyst recommendation revisions. In particular, we find that analysts tend to keep past good "accurate'' recommendations for too long, and to drop past bad "inaccurate'' recommendations too quickly. That is, incentives to maintain a 'good' reputation---a good-looking recommendation list in front of customers---appear to aggravate rather than alleviate conflicts of interest.
We find that recommendation revisions primarily reflect retrospective considerations, i.e., the performance of analysts' outstanding recommendations (i.e., the consistency of the recommendation with the return since the recommendation was made), rather than prospective considerations about future returns. The impact of performance on recommendations is significantly time varying: initially both very poor and very good performance raise the hazards of downward revisions, but as the duration of a recommendation increases, cumulative returns only have monotonic negative effects on the hazards of recommendation downgrades. These findings highlight the inappropriateness of both single-period classification models such as probits, and of proportional hazard models that only account for duration dependence via a time-varying baseline hazard. More generally, we exhaustively characterize the impacts of covariates on recommendation changes, for example, showing that the information content of earnings announcements is fully incorporated into recommendations within one week, and that the likelihood of revisions falls sharply with experience in an analyst's first three years (the odds of a revision in an analyst's first year are roughly three times that of an analyst with three years experience), but then plateaus.
We then ask: how does a recommendation's past performance influence the sensitivity of recommendations to future returns? Overall, analysts appear prescient: downgrades (upgrades) are more likely when future returns are bad (good). However, past poor performance of a recommendation, e.g., past very negative returns following a strong buy, dramatically reduce the sensitivity of recommendations to future returns. In particular, if since a strong buy or buy recommendation was issued, returns were highly negative, then downgrades are very likely, independently of whether future returns are low or high. In contrast when past returns were higher, the likelihood of a recommendation revision becomes much more sensitive to future returns, implying that an analyst's recommendation decision (the maintainence of a strong buy versus a downgrade) contains more information about future returns when past returns have been higher. The sensitivity of upgrades from buy and hold recommendations to future returns is more subtle, as it is time varying. Very negative past returns especially for longer-held recommendations mean that the probability of a revision is roughly orthogonal to future returns; while qualitatively, upgrade probabilities are most sensitive to future returns when past returns have not been too high.
Contagious Default
Driven by the corporate default data for U.S. public firms, I develop a method to estimate corporate default intensities that can incorporate both contagious defaults---defaults correlated through the network of contractual or financial relationships---and informatively-censored events such as mergers and acquisitions---an option to avoid bankruptcy for distressed firms. In my semiparametric hazard model, default correlations among firms and through time are modeled by two-dimensional penalized tensor-product splines. Dropouts caused by mergers and acquisitions, are treated as competing risks that are correlated with corporate defaults through a latent random variable on which they have different loadings. My approach can also be interpreted based on a mutually exciting multivariate point process model, where the direct impact of one firm's default on other firms is quantified by the financial distance and where defaults from external sources are modeled in the dependence of observable covariates and unobservable frailties. My results show that the proposal model is able to estimate the correlated corporate defaults in a more realistic way and predict the firms' credit risks with higher confidence level than the dynamic frailty model of Duffie et al. (2009), which implicitly accounts for excessive correlation caused by contagion and informatively-censored events via a common unobservable macroeconomics risk factor.
Is Cash Flow Really King? Earnings Quality and Stock Returns (September 2008)
Past research indicates that net operating cash flows are more persistent than accruals (the difference between accounting earnings and operating cash flows); firms with high operating accruals earn lower average returns than firms with low operating accruals. More recent research finds that once cash flows are accounted for, the negative association between accruals and future abnormal returns vanishes. In this paper, I analyze the complex relation between accruals, operating cash flow and subsequent returns using semiparametric and nonparametric regression approaches. I uncover a highly nonlinear and non-monotonic relationship between accruals and operating cash flows, indicating that one cannot just incorporate the impact of cash flows on abnormal future returns and ignore accruals. In particular, in the long run, accruals contain more information about future abnormal returns, while in the short run, cash flow "is king'', and has more predictive power about future abnormal returns. My findings suggest that the assumptions made by previous researchers are not valid, and that their analysis cannot appropriately capture the underlying relations, and can give rise to misleading inference. I also attempt to investigate the underlying mechanism of the accruals anomaly by determining whether and how sophisticated information intermediaries such as financial analysts use the information contained in accruals and cash flows in their earnings forecasting process. I find that that analysts are over-optimistic about firms reporting high accruals and do not correctly assess the quality of earnings. These results favor a behavioral explanation for accrual anomaly.
WORK IN PROGRESS:
Liquidity Dynamics and Stock Returns: a Bayesian Nonparametric Approach
This paper explores the joint dynamics of liquidity risks, default risks, volatility, and stock returns using a gaussian process dynamical model. Employing the flexibility of this statistical model that accommodates nonlinear and nonstationary time series, I comprehensively investigate the co-movements and interrelations of liquidity, default, volatility and returns. The results demonstrate that the intertemporal associations between those variables are not stationary, and that they vary predictably over time and with different market conditions.
After Filing Chapter 11
I investigate the hazards of mergers, emergences and liquidations after firms file Chapter 11 bankruptcy using a quantile survival analysis model. Applying this semiparametric duration model, I find considerable differences in the determinants of mergers, emergences and liquidations. Such differences reflect the primary cause for the bankruptcy---whether it is triggered by operational risks, funding liquidity risks, or market liquidity risks, etc.. Different investment values could be assigned according to these characteristics of the firms at the time when they file Chapter 11 bankruptcy. These findings shed light on the nature of optimal trading strategies of distressed stocks.
Multi-Period Corporate Default Prediction Using Joint Model of Accelerated Failure Time and Longitudinal Data
In this paper, in order to take into account the time-varying default hazard and correlation between the bankruptcies, I explore the joint modeling approach under accelerate failure time assumption by maximizing the joint likelihood function of the bankruptcy time and longitudinal data with random effects (frailty). This model produces more accurate out-of-sample forecasts than alternative models by incorporating the dynamics of both observed and unobserved (frailty) covariates. A Monte Carlo EM algorithm is used to estimate all unknown parameters.