Working papers
Duration Bias
When aggregate performance statistics are computed by counting each entity once from an accumulated historical record, they can substantially misrepresent the population that investors actually face at any point in time. The databases themselves are accurate; the bias lies in how they are used. I formalize this duration bias and prove that it arises whenever entity lifetime is correlated with outcomes. The most prominent example is Bessembinder's (2018) nding that 57% of U.S. stocks underperform Treasury bills. This statistic describes the historical archive, not the market investors face. The appropriate approach is to compute the cross-sectional underperformance rate at each point in time, among stocks actually available to investors, and average across time. This observation-weighted rate is 34.5%, a gap of 23 percentage points. The result holds at every investment horizon: even at one year, the database rate overstates the observation-weighted rate by 12 percentage points (t = 5.93). A purely backward-looking version using zero future information con rms the pattern. The gap arises because short-lived, volatile stocks cycle through the database far faster than they occupy the market, in ating their weight in the archive. The same duration bias in inates measured failure rates in the CRSP Mutual Fund Database by 10 percentage points, even though it is explicitly designed to be survivorship-bias-free. International evidence from 50 countries con rms the pattern. Using a standard portfolio choice framework, the misperception costs on the order of $200,000 per household over a 30-year horizon. I also develop an observation-weighted variance decomposition showing that standard bootstrap simulations on unbalanced panels create synthetic entities whose variance properties match no actual security.
When aggregate performance statistics are computed by counting each entity once from an accumulated historical record, they can substantially misrepresent the population that investors actually face at any point in time. The databases themselves are accurate; the bias lies in how they are used. I formalize this duration bias and prove that it arises whenever entity lifetime is correlated with outcomes. The most prominent example is Bessembinder's (2018) nding that 57% of U.S. stocks underperform Treasury bills. This statistic describes the historical archive, not the market investors face. The appropriate approach is to compute the cross-sectional underperformance rate at each point in time, among stocks actually available to investors, and average across time. This observation-weighted rate is 34.5%, a gap of 23 percentage points. The result holds at every investment horizon: even at one year, the database rate overstates the observation-weighted rate by 12 percentage points (t = 5.93). A purely backward-looking version using zero future information con rms the pattern. The gap arises because short-lived, volatile stocks cycle through the database far faster than they occupy the market, in ating their weight in the archive. The same duration bias in inates measured failure rates in the CRSP Mutual Fund Database by 10 percentage points, even though it is explicitly designed to be survivorship-bias-free. International evidence from 50 countries con rms the pattern. Using a standard portfolio choice framework, the misperception costs on the order of $200,000 per household over a 30-year horizon. I also develop an observation-weighted variance decomposition showing that standard bootstrap simulations on unbalanced panels create synthetic entities whose variance properties match no actual security.
The Anomaly Zoo Has Many Alphas but Few Dimensions with George Chalamandaris, and Nikolaos Topaloglou
Do stock market anomalies make investors better off? We test whether any of 153 anomalies improves the return distribution available to risk-averse investors using a joint spanning test based on second-order stochastic dominance. The anomaly zoo is economically redundant: the largest improvement in bad-state outcomes is 1.3 basis points per week—roughly one-tenth of a typical significant alpha—and statistically insignificant. The test has power: reducing the benchmark from eight to seven factors causes spanning to collapse. Cross-sectional evidence shows that tail improvements and pricing errors co-move. The anomaly zoo contains many alphas but few new dimensions of downside protection.
Do stock market anomalies make investors better off? We test whether any of 153 anomalies improves the return distribution available to risk-averse investors using a joint spanning test based on second-order stochastic dominance. The anomaly zoo is economically redundant: the largest improvement in bad-state outcomes is 1.3 basis points per week—roughly one-tenth of a typical significant alpha—and statistically insignificant. The test has power: reducing the benchmark from eight to seven factors causes spanning to collapse. Cross-sectional evidence shows that tail improvements and pricing errors co-move. The anomaly zoo contains many alphas but few new dimensions of downside protection.
Open-Source Activities and Stock Returns: Evidence from GitHub Engagement with Cagri Berk Onuk and Yunfei Zhao
We examine whether GitHub engagement predicts stock returns. Using data linking S&P 1500 firms to GitHub activity (2010–2022), we construct a market-equity scaled PCA measure of engagement based on GitHub stars, forks, watchers, and repositories. Portfolios sorted on this composite measure yield significant abnormal returns: the high-minus-low spread is 1.45% monthly, compounding to 52.48% over five years, with robust Fama-French and q-factor alphas. Predictability strengthens for hard-to-fake metrics and survives fake-star cleansing. GitHub engagement predicts returns at intermediate horizons (peaking at a five-month lag) but not at short horizons, contrasting with traditional innovation proxies that predict at immediate lags, consistent with a production-side signal that materializes through integration, deployment, and adoption cycles. A decomposition shows GitHub stars reflect future profitability and intangible value. These findings suggest open-source engagement signals organization capital that investors underweight, contributing to literatures on alternative data and intangible assets.
We examine whether GitHub engagement predicts stock returns. Using data linking S&P 1500 firms to GitHub activity (2010–2022), we construct a market-equity scaled PCA measure of engagement based on GitHub stars, forks, watchers, and repositories. Portfolios sorted on this composite measure yield significant abnormal returns: the high-minus-low spread is 1.45% monthly, compounding to 52.48% over five years, with robust Fama-French and q-factor alphas. Predictability strengthens for hard-to-fake metrics and survives fake-star cleansing. GitHub engagement predicts returns at intermediate horizons (peaking at a five-month lag) but not at short horizons, contrasting with traditional innovation proxies that predict at immediate lags, consistent with a production-side signal that materializes through integration, deployment, and adoption cycles. A decomposition shows GitHub stars reflect future profitability and intangible value. These findings suggest open-source engagement signals organization capital that investors underweight, contributing to literatures on alternative data and intangible assets.
Wrapped Stablecoins and Systemic Risk in DeFi with Pairie Koh
Wrapped stablecoins convert base stablecoins into receipt tokens that can be pledged repeatedly across DeFi lending protocols, implementing a form of shadow banking on-chain. We develop a model in which this recursive collateral reuse generates an endogenous liquidity multiplier that expands leverage in normal times but amplifies liquidation cascades when collateral constraints tighten. Using novel on-chain data from Aave deployments across eight blockchain networks, we document three findings. First, using regulatory announcements as instruments, we show that wrapped stablecoin creation causally increases recursive leverage: a one-standard-deviation increase raises the liquidity multiplier by 7\% relative to its mean. Second, higher pre-event recursive leverage amplifies cryptocurrency volatility during stress episodes, with economically large effects concentrated in equally-weighted rather than value-weighted measures. Third, cross-chain bridge connectivity, not chain size, determines contagion speed: doubling bridge connections reduces time to stress arrival by 67\% during the March 2023 USDC depeg. Wallet-level evidence confirms that agents execute the deposit-borrow-redeposit loops predicted by our model. Our results highlight a fundamental tradeoff: deeper collateral reuse improves funding efficiency but increases exposure to systemic liquidation cascades.
Wrapped stablecoins convert base stablecoins into receipt tokens that can be pledged repeatedly across DeFi lending protocols, implementing a form of shadow banking on-chain. We develop a model in which this recursive collateral reuse generates an endogenous liquidity multiplier that expands leverage in normal times but amplifies liquidation cascades when collateral constraints tighten. Using novel on-chain data from Aave deployments across eight blockchain networks, we document three findings. First, using regulatory announcements as instruments, we show that wrapped stablecoin creation causally increases recursive leverage: a one-standard-deviation increase raises the liquidity multiplier by 7\% relative to its mean. Second, higher pre-event recursive leverage amplifies cryptocurrency volatility during stress episodes, with economically large effects concentrated in equally-weighted rather than value-weighted measures. Third, cross-chain bridge connectivity, not chain size, determines contagion speed: doubling bridge connections reduces time to stress arrival by 67\% during the March 2023 USDC depeg. Wallet-level evidence confirms that agents execute the deposit-borrow-redeposit loops predicted by our model. Our results highlight a fundamental tradeoff: deeper collateral reuse improves funding efficiency but increases exposure to systemic liquidation cascades.
Agnostic Stochastic Discount Factors in Small Samples: Theory, Momentum, and Cross-Market Evidence with Richard Roll and Junbo Wang
- The first version with Richard Roll, "An Agnostic and Practically Useful Estimator of the Stochastic Discount Factor"
- The revised version with Richard Roll and Junbo Wang, "An Agnostic and Practically Useful Estimator of the Stochastic Discount Factor"
- New York University, Louisiana State University, Texas Tech, World Finance 2021, LSU (2021), MFA 2022
- The 2nd prize, best paper award, The 14th SME International Conference
Climate Risk Preparedness and the Cross Section with Siddhartha Chib, Leifei Lyu, and Khaled Obaid
- FinTech HK Conference 2024
- What we found is distinct from climate risk exposure!!
Forecasting the Term Structure of Equity Betas: Implications for Valuation with Niklas Augustin
Estimating firm discount rates remains a fundamental challenge in asset pricing and corporate finance. We propose a novel approach that forecasts discrete elements of the beta term structure. Our results show that firm betas exhibit predictable variation across horizons and can be reliably forecasted for at least ten years. Incorporating expected betas into factor models reveals substantial dispersion in costs of equity, with an annual difference of approximately five percentage points between the top and bottom deciles of firms. Adjusting for these forward-looking costs of equity in discounted cash flow models reduces pricing errors by 10\% relative to standard methods. Furthermore, accounting for changes in expected betas is necessary to construct truly market-neutral long-term portfolios. These findings highlight the importance of horizon-specific beta forecasts for improving the estimation of discount rates, valuation, and investment decisions.
Estimating firm discount rates remains a fundamental challenge in asset pricing and corporate finance. We propose a novel approach that forecasts discrete elements of the beta term structure. Our results show that firm betas exhibit predictable variation across horizons and can be reliably forecasted for at least ten years. Incorporating expected betas into factor models reveals substantial dispersion in costs of equity, with an annual difference of approximately five percentage points between the top and bottom deciles of firms. Adjusting for these forward-looking costs of equity in discounted cash flow models reduces pricing errors by 10\% relative to standard methods. Furthermore, accounting for changes in expected betas is necessary to construct truly market-neutral long-term portfolios. These findings highlight the importance of horizon-specific beta forecasts for improving the estimation of discount rates, valuation, and investment decisions.
- The Third Crowell Prize Winner, 2025, PanAgora Asset Management; direct download here
- University of Massachusette, Lowell March 2026, Technical University of Munich (TUM) School of Management, 2nd Workshop on Advances in NLP and Generative AI in Finance and Management, Munich, Germany, May 19-20, 2025, University of Missouri Columbia, Missouri State University, University of Missouri St. Louis, the Fields Institute for Research in Mathematical Sciences at University of Toronto 2022, Poznan University, Blackrock 2023, FinTech City University of Hong Kong 2023, Northern Finance Association meeting 2023, Depaul University in November 2025, MFA 2026,
The Magnificent Ten Equity Factor Model with Argyro Kofina, Ioannis Psaradellis, and Nikolas Topaloglou
- A ten-factor SSD-based sparse model remains optimal and consistently outperforms all leading benchmarks.
- The concept based on stochastic dominance considering all welfare.
Changing Expected Returns Can Induce Non-zero but Unprofitable (Illusory) Return Autocorrelation with Richard Roll and Avanidhar Subrahmanyam
- 6th JAAF India Symposium (January 2022), Auckland University of Technology, The 14th SME International Conference
Transmission Bias in Financial News with Khaled Obaid We examine how financial news is distorted as it spreads across different news outlets, akin to the “telephone game”. Using a sample of exclusive articles from the Wall Street Journal, we use ChatGPT-4 to quantify the distortion introduced in the information environment as competing news outlets retell these exclusive stories. We find strong evidence that retelling articles tend to be more opinionated and negative and less factual and appealing compared to the original story. Factors that influence the distortion include whether the stories are retold by news outlets that are less specialized, the time-lapse between the original story and its retelling, and the presence of competing narratives from other news outlets. Our findings indicate that media distortion affects asset prices, influences trading behaviors, and leads to disagreement among analysts.
Major Revision
Animating Stock Markets by Tomasz Kaczmarek and Kuntara Pukthuanthong
- Even wonder if we can use graphs to predict graphs in a form of animation. You got it. We can! Coming soon.
- Finalist 2024, Crowell Prize, PanAgora Asset Management
- MFA 2024; Future Finance Info 2024
Cracking the Code: The Networking Matrix of Finance Academia with Sujiao (Emma) Zhao
- Bank of Portugal (July 2023); FMA (2023)
Average and Marginal Tobin’s q as Indicators of Future Growth Opportunities, Expected Return, and Risk with Richard Roll
Contrary to popular opinion, average Tobin’s q is a better indicator of future growth opportunities than marginal Tobin’s q. We derive a curious relation between average and marginal q: the more profitable a new investment opportunity, the smaller will be the increase in average q when the opportunity is undertaken. Average q is inversely related to the cost of equity capital, so it represents an inverse measure of risk. The closely- related book/market ratio is also a measure of risk in the cross-section.
Contrary to popular opinion, average Tobin’s q is a better indicator of future growth opportunities than marginal Tobin’s q. We derive a curious relation between average and marginal q: the more profitable a new investment opportunity, the smaller will be the increase in average q when the opportunity is undertaken. Average q is inversely related to the cost of equity capital, so it represents an inverse measure of risk. The closely- related book/market ratio is also a measure of risk in the cross-section.
Rational Apathy: Unveiling the Hidden Consequences of Workplace Automation with Zhihua Wei, Aoran Zhang, and Yunfei Zhao
- FMA (2022)
Foreign Exchange Exposure – No Returns, only risk with Heikki Lehkonen and Khaled Obaid
We use generative AI to capture the foreign exposure. Returns of portfolios formed based on US dollar foreign exchange exposure are found to have a negative relationship with the absolute value of the foreign exchange exposure. The result is state-dependent and visible during the appreciation periods of the US dollar. We relate this effect to mispricing where investors fail to consider to true impacts of the exchange rate changes on firm values and report several results supporting this claim.
We use generative AI to capture the foreign exposure. Returns of portfolios formed based on US dollar foreign exchange exposure are found to have a negative relationship with the absolute value of the foreign exchange exposure. The result is state-dependent and visible during the appreciation periods of the US dollar. We relate this effect to mispricing where investors fail to consider to true impacts of the exchange rate changes on firm values and report several results supporting this claim.
Permanent Working Papers
A Generalized Machine Learning Framework for Linear Factor Model Test with Christopher Jones, Jinchi Lv and Junbo Wang
We introduce a generalized statistical learning method, sparse orthogonal factor regression (SOFAR), in testing linear factor models with both large numbers of factors and testing assets. Our approach encompasses most of the existing methods in the literature and can be used in many other scenarios with large data sets. Applying SOFAR, we can select the PC factors from the whole swath of 219 candidate factors proposed by the literature simultaneously, identify test assets associated with the selected PC factors, and interpret them. We can also select the PC factors and correlated characteristics in the IPCA framework without bootstrapping. Without firm characteristics instrumenting, we find that four PC factors (market, investment, intangible, and frictions) are relevant to the covariance of asset returns and three types of factors (profitability, asset liquidity, and liquidity bets) price assets in cross-section. With characteristics as instruments, we only identify one factor, and the correlated characteristics are beta, size, momentum, and liquidity.
- LSU (2020); SOFIE UCSD 2021; NBER-NSF 2021
We introduce a generalized statistical learning method, sparse orthogonal factor regression (SOFAR), in testing linear factor models with both large numbers of factors and testing assets. Our approach encompasses most of the existing methods in the literature and can be used in many other scenarios with large data sets. Applying SOFAR, we can select the PC factors from the whole swath of 219 candidate factors proposed by the literature simultaneously, identify test assets associated with the selected PC factors, and interpret them. We can also select the PC factors and correlated characteristics in the IPCA framework without bootstrapping. Without firm characteristics instrumenting, we find that four PC factors (market, investment, intangible, and frictions) are relevant to the covariance of asset returns and three types of factors (profitability, asset liquidity, and liquidity bets) price assets in cross-section. With characteristics as instruments, we only identify one factor, and the correlated characteristics are beta, size, momentum, and liquidity.
AI Narrative and Stock Mispricing with Arka Bandyopadhyay and Dat Mai
- https://alphaarchitect.com/2023/06/ai-exposures/
- https://www.barrons.com/articles/ai-nvidia-tech-winners-has-beens-e0dec10
Diversity Narrative and Equity in Firm Leadership with Arka Bandyopadhyay and Dat Mai
We provide causal evidence that the narrative of diversity from the New York Times articles has nudged the corporations to choose female CEOs to be equitable in terms of gender of firm leadership. This channel is independent of the board gender diversity, which was mandated in 2017 in California and later repealed in 2022. Surprisingly, the diversity narrative channel fails to explain the election of Indian CEOs in several HiTech firms over the last few years. We argue that the election of Indian CEOs was motivated to create a favorable image of Tech firms to Indian and Chinese labor. We conclude that providing equity in firm leadership in terms of ethnic diversity is harder compared to gender diversity.
We provide causal evidence that the narrative of diversity from the New York Times articles has nudged the corporations to choose female CEOs to be equitable in terms of gender of firm leadership. This channel is independent of the board gender diversity, which was mandated in 2017 in California and later repealed in 2022. Surprisingly, the diversity narrative channel fails to explain the election of Indian CEOs in several HiTech firms over the last few years. We argue that the election of Indian CEOs was motivated to create a favorable image of Tech firms to Indian and Chinese labor. We conclude that providing equity in firm leadership in terms of ethnic diversity is harder compared to gender diversity.
Interest Rates and Real Estate Values: The Divergence Effects of Real Rates and Expected Inflation with Richard Roll
A common belief is that higher interest rates reduce real estate values, but corroborating evidence is underwhelming. A sensible explanation suggests heterogeneity across countries is that mortgage interest deductibility for the U.S. personal income tax makes much of the real value of the mortgage principal tax deductible when inflation is high. This implies that nominal mortgage rates could positively affect U.S. house prices. We find evidence consistent with this possibility; the inflation component of nominal interest is associated positively with U.S. real estate prices but negatively with those in Canada, a country without mortgage interest deductibility.
A common belief is that higher interest rates reduce real estate values, but corroborating evidence is underwhelming. A sensible explanation suggests heterogeneity across countries is that mortgage interest deductibility for the U.S. personal income tax makes much of the real value of the mortgage principal tax deductible when inflation is high. This implies that nominal mortgage rates could positively affect U.S. house prices. We find evidence consistent with this possibility; the inflation component of nominal interest is associated positively with U.S. real estate prices but negatively with those in Canada, a country without mortgage interest deductibility.
An Unambiguous Statement of Interest Rate Parity with Lee R. Thomas III
Interest rate parity (IRP) states that the difference between the interest rates of two currencies equals the expected percentage change in the associated exchange rate. Unfortunately, the IRP is ambiguous. There are two, inconsistent values of the expected percentage change in the exchange rate, depending on which currency is treated as the numeraire. We derive a corrected IRP condition that is unambiguous but subtly different. The equilibrium interest rate differential equals the percentage difference between the spot exchange rate and the geometric mean of the future exchange rate distribution, not its expected value.
Interest rate parity (IRP) states that the difference between the interest rates of two currencies equals the expected percentage change in the associated exchange rate. Unfortunately, the IRP is ambiguous. There are two, inconsistent values of the expected percentage change in the exchange rate, depending on which currency is treated as the numeraire. We derive a corrected IRP condition that is unambiguous but subtly different. The equilibrium interest rate differential equals the percentage difference between the spot exchange rate and the geometric mean of the future exchange rate distribution, not its expected value.