Journal of Investment Strategies

Ali Hirsa
Professor, Columbia University & Managing Partner, Sauma Capital LLC

Welcome to the third issue of the fourteenth volume of The Journal of Investment Strategies, which contains three research papers.

In the first paper in this issue, “A new approach to asset pricing models: the term structure of leverage and refinancing risk”, Esra Karpuz Demir and Guven Sevil study how financial leverage relates to equity returns, with particular emphasis on whether investors price leverage differently depending on a firm’s debt maturity structure and debt refinancing intensity. Rather than treating leverage as a single aggregate measure, Karpuz Demir and Sevil construct portfolios based on shortterm versus long-term leverage and refinancing risk and then evaluate the portfolios’ return behavior using multiple asset-pricing factor models. The core question is whether shareholders assign varying levels of importance to firms’ debt maturity profiles and, consequently, to financial leverage derived from short- and long-term debt. The authors’ empirical results suggest that higher total leverage is associated with lower expected stock returns (or, in some specifications, has no meaningful effect), while greater short-term borrowing is associated with higher expected returns (a statistically significant increase of about 2%). This indicates that investors demand compensation specifically for risks tied to short-term leverage and refinancing, consistent with the view that leverage’s effect on valuation depends on its maturity composition, not merely its level. The study also compares factor models’ ability to explain returns, finding that performance varies with portfolio construction and with whether portfolios are equally weighted or value weighted. Among the tested combinations, the Fama–French six-factor model produces the lowest pricing error for the equally weighted risk level equalized volatility (RLEV) portfolio, and such portfolios generally exhibit stronger risk-adjusted performance compared with the market across standard metrics (eg, the Sharpe ratio, the Modigliani–Modigliani measure and the Treynor ratio). Overall, the authors highlight the relevance of the leverage maturity structure and refinancing intensity for understanding leverage-related risk premiums, noting that future work could extend the data set and incorporate additional risk factors for a more comprehensive analysis.

In “On profitability and maximum tolerable latency in the high-frequency trading of a microtrend anomaly”, the issue’s second paper, James A. Primbs, Bogdan Mukhametkaliev, B. Ross Barmish and Sean Warnick empirically examine whether the single-stock Sieczka–Hołyst microtrend anomaly can be profitably exploited using the slippage-free strategy developed by three of the authors in 2022.¹ Using Nasdaq ITCH data for the 31 Dow Jones Industrial Average (DJIA) component stocks from 2018, they find that an idealized zero-latency trader could have earned up to around 77 basis points per day on an equally weighted DJIA-31 portfolio, and above 3% per day for certain individual stocks. Because the strategy “picks off” favorable bid/ask quotes in the direction of a microtrend, it resembles latency arbitrage, where profitability is highly sensitive to speed. To quantify this, the authors estimate the strategy’s maximum tolerable latency (MTL), defined as the largest round-trip latency under which the strategy remains profitable. For the equally weighted DJIA-31 portfolio in 2018, average MTL is 14.6 (units as reported in the paper), while individual stocks show average MTLs ranging from roughly 0 to 40. On some days, however, MTLs are much larger (values above 50 are not uncommon, and extremes can exceed 1000), suggesting that exploitability may still be feasible given Nasdaq’s advertised sub-50 colocation latencies. Overall, the results of the study characterize the effective “speed” and exploitability of DJIA stocks on Nasdaq in 2018. The authors also run regressions linking zero-latency returns and MTL to market features, finding significant associations with average microtrend length, the probability of slippage-free opportunities, bid–ask spreads and the average time between executions. Future work could test enhanced order types (eg, immediate or cancel, fill or kill), incorporate fees and rebates into profitability and MTL, and use more sophisticated statistical or machine learning methods to predict the best stocks and days to target and to improve the trading rules beyond the simple strategy studied in this paper. However, accurate stock trend forecasting is widely viewed as challenging in financial economics, given the nonlinear and interdependent nature of market dynamics. In many settings, traditional statistical and machine learning models may have difficulty capturing complex temporal dependencies and interstock relationships.

Our final paper, “All models are wrong. Some might be OK!” by Laura Ryan, Baoqing Gan and Geoffrey J. Warren, examines three primary types of model uncertainty: aleatoric uncertainty (inherent randomness); epistemic uncertainty (due to parameter and variable selection); and functional form uncertainty (ie, relating to the mathematical structure of the relationship). While all three types pose challenges, the authors argue that functional form uncertainty is uniquely problematic: the “true” data-generating process is rarely known, and ambiguity often persists despite large data sets or advanced modeling tools. To illustrate these challenges, Ryan et al use equity–bond correlation modeling as a case study. They conduct an experiment using simulated data in which the true correlation is governed by a thresholdbased data-generating process. They then evaluate whether standard and advanced techniques – ranging from ordinary least squares regression and bootstrapping to density-based spatial clustering of applications with noise (DBSCAN) and Gaussian mixture models – can successfully recover this underlying structure. Their results show that when functional form uncertainty is ignored, standard methods frequently misestimate correlations. Even sophisticated machine learning methods struggle to detect the true structure unless the noise in the data is unrealistically low. These difficulties are further magnified in predictive contexts, where structural shifts and overfitting can rapidly degrade model performance. Ryan et al conclude that practitioners should move away from the pursuit of a single “correct” model and instead embrace uncertainty through ensemble-based or probabilistic approaches. By maintaining transparency regarding model limitations, their paper provides a roadmap for more resilient financial modeling and portfolio construction.

The editorial board and I extend our sincere gratitude to you, our valued readers, for your unwavering support and interest in our journal. We are delighted to present an expanding collection of practical papers on diverse topics related to modern investment strategies, contributed by both academic and industry experts.