Francisco López Herrera

When a machine learning trading strategy generates positive risk-adjusted returns, how much of the performance is attributable to the forecasting model and how much to the execution mechanism that surrounds it? We propose a Monte Carlo signal-vs-mechanism decomposition that holds a dual-buffer straddle execution protocol fixed while varying only the directional signal between three ensemble learners (XGBoost, Random Forest, Gradient Boosting) and 1,000 random draws per security at four fill-rate scenarios. Implementing the decomposition on 48 U.S.-listed securities (2018–2025) through expanding-window walk-forward validation produces 216,864 out-of-sample predictions across 72 sequential windows. Under perfect fills, ML and random signals yield statistically indistinguishable Sharpe ratios; as fills degrade, ML amplification reaches 3.7×, 2.8×, and 1.8× across high-, normal-, and low-volatility regimes at 60% fill. The underlying accuracy gradient (81.5% vs. 60.8%; Cohen's d = 5.36) aligns monotonically with abnormal trading volume and VIX, validating a behavioral interpretation. The framework generalizes to any ML application where signal and mechanism are empirically entangled.

We extend the buffer-based limit order execution framework for machine learning trading strategies by introducing a dual-buffer mechanism that applies patient limit order exits to both legs of a straddle position. Building upon the single-buffer "patient trader" approach—where only the losing leg exits via limit orders—our enhancement captures additional profit from the winning position through symmetric execution at favorable prices.Testing three ML algorithms (XGBoost, Random Forest, Gradient Boosting) across 48 global securities over 1,989 trading days (2018–2025) with buffer levels from 10% to 40%, we find the dual-buffer approach dramatically outperforms the single-buffer methodology. At the 30% buffer level, the dual-buffer strategy achieves Sharpe ratios of 17.40 under full execution versus 3.17 for the original single-buffer approach. A fill rate degradation analysis—evaluating performance as the kept leg’s buffer execution probability decreases from 100% to 50%—demonstrates graceful degradation: at the 30% buffer, Sharpe ratios decline from 17.40 to 10.29, still exceeding the original methodology by a factor of three and buy-and-hold by a factor of seventeen. The original single-buffer results constitute the performance floor, achieved when the kept leg always closes at market. Comprehensive risk analysis using VaR, CVaR, and Maximum Drawdown confirms enhanced risk-adjusted performance across all fill rate assumptions.Robustness analyses across five subperiods—including the COVID-19 crisis—demonstrate strategy stability, with Deflated Sharpe Ratios confirming significance after adjusting for 576 strategy configurations. Our findings demonstrate that the dual-buffer mechanism generates substantial execution alpha under full buffer execution independent of prediction quality, while ML prediction’s directional accuracy becomes the critical floor-preserving mechanism as kept-leg fill rates degrade.

Machine learning's value for liquidity prediction is asymmetric: substantial for identifying opportunities, minimal for detecting crises. We develop a framework predicting next-day liquidity regime transitions (Improving, Stable, Deteriorating) for 47 global securities across 12 countries from 2018 to 2025, benchmarking four ML models against five traditional approaches including mean-reversion and Markov transition. Random Forest achieves 62.5% accuracy versus 59.1% for mean-reversion (p<0.001, 97.9% win rate), but class-level analysis reveals the source of this advantage: ML excels at predicting liquidity improvements (+4.1 percentage points F1-score over mean-reversion) while providing minimal gains for deterioration (+0.6 percentage points). Feature importance analysis shows that volume surges and volatility declines—signals beyond simple mean-reversion—drive ML's advantage for improvement prediction. For practitioners, we recommend a hybrid approach: ML models for opportunity identification, simple mean-reversion heuristics for risk monitoring. Economic value analysis confirms ML generates $102,000 in annual advantage at baseline trade sizes.

We introduce a buffer-based limit order execution framework for machine learning trading strategies. Our "straddle-then-reveal" approach operates on a two-day cycle: on day t, the trader enters simultaneous long and short positions at market open (Open_t); at day t's close, ML models—using features computed from historical data through day t—predict whether Close_{t+1} > Open_t; on day t+1, the "wrong" leg exits via patient limit orders placed at a buffer percentage from the intraday price extreme. Notably, anchoring the target to Open_t rather than Close_t proves essential: the alternative yields accuracy indistinguishable from random guessing, suggesting the intraday price path embeds predictive information about next-day movement. Testing three ML algorithms with standard and quantum-inspired feature selection across 48 global securities over 1,989 trading days (2018–2025), we find striking results. Models achieve classification accuracy near 70%, and the buffer mechanism translates this predictive power into realized profits: at the 10% buffer, average Sharpe ratios reach 14.86 with 85.5% win rates; at the conservative 30% buffer, Sharpe ratios remain strong at 7.26 with 65.6% win rates. The Sortino-to-Sharpe ratio ranges from 2.2× to 7.5× across configurations, confirming favorable return asymmetry. Robustness tests across five subperiods—including the COVID-19 crisis—confirm strategy stability. Slippage sensitivity analysis demonstrates profitability persists across realistic transaction cost scenarios. Our findings suggest that execution mechanism design can be as important as prediction accuracy for trading strategy performance.

El propósito de este libro es mostrar los principales replanteamientos de los con-ceptos centrales de la Teoría Moderna del Portafolio, motivados por las limitaciones de esta teoría para enfrentar la realidad. Por tal motivo, en esta obra se hace un breve recorrido sobre el desarrollo de la teoría para el análisis de inversiones hasta llegar a la Teoría Postmoderna del Portafolio.1952 fue el año en el que surgen las ideas centrales de la teoría financiera. En ese año, Harry Markowitz publica su artículo Selección de Portafolios y en él sienta las bases de la llamada Teoría Moderna del Portafolio, pero en el mismo año Andrew Roy también publica un artículo La Seguridad Primero y la Tenencia de Activos en donde muestra las ideas principales de la Teoría Postmoderna del Portafolio como alternativa para el análisis y selección de inversiones. En esta obra se ofrece un panorama general del avance de la teoría de inversiones, y la aplicación de algunas de sus principales métricas, a la luz de la Teoría Postmoderna del Portafolio.

This study comparatively evaluated the forecasting performance of a constant beta and two time-varying beta process specifications. Returns for 23 stocks were forecasted for several horizons in 2019–2020. The autoregressive and random walk betas showed superior forecasting performance before and during the COVID-19 pandemic, respectively. In a Diebold–Mariano test, the constant beta specification never dominated both time-varying beta models. Beta specification type should be considered when forecasting based on capital asset pricing or market models, particularly during crises. Results of rolling regressions based on arbitrage pricing theory considering other risk factors suggest time-varying systematic risk factors beyond market risks.

El objetivo de este documento es analizar la sincronización de los ciclos económicos de México y Estados Unidos durante el periodo 1981-2017 mediante la estimación de un coeficiente de correlación condicional dinámica que permite tener una estimación para cada periodo de tiempo. Los resultados, obtenidos a partir de distintos indicadores de producción y métodos de eliminación de tendencia, muestran un aumento desde la apertura de la economía mexicana a mediados de la década de 1980, especialmente durante las recesiones de 2001-2002 y 2008-2009 y también una serie de descensos aislados, explicados por diferencias en los ritmos de crecimiento de ambas economías, y una declinación sostenida en la fase pos-Gran Recesión que se explica principalmente por reducciones en el comercio exterior. SYNCHRONIZATION OF THE BUSINESS CYCLES OF MEXICO AND THE UNITED STATES: A DYNAMIC CORRELATION APPROACHABSTRACTThe objective of this paper is to analyze the business cycle synchronization of Mexico and the United States over the period 1981-2017 by estimating a dynamic conditional correlation coefficient that allows us to have an estimate for each time period. The results, obtained from different production indicators and different de-trending methods, show an increase in this synchronization after the opening of the Mexican economy in the mid-eighties, especially during the common recessions of 2001-2002 and 2008-2009, and some isolated drops explained by differences in the growth rates of both economies as well as a sustained decline in the post-Great Recession phase resulting from the decline of international trade.

Para entender el riesgo bursátil durante las crisis financieras internacionales de 1998 a 2021, se analizó la volatilidad variante de los siete índices sectoriales de la Bolsa Mexicana de Valores de 1998 a 2021 mediante el modelo de heteroscedasticidad condicional auto regresiva asimétrica (GJR-GARCH). Los resultados fueron acordes con las hipótesis de trabajo pues se encontró que las principales crisis internacionales tuvieron efectos diferenciados. En la crisis rusa de 1998 el sector industrial tuvo el mayor riesgo; en las crisis latinoamericanas de 2000 a 2002 y la dot-com, los sectores salud, servicios financieros y telecomunicaciones presentaron los mayores niveles de desviación estándar; de 2004 a 2008 cinco sectores mostraron alta volatilidad, sobresaliendo el sector industrial y durante la reciente crisis de COVID-19 los sectores más riesgosos fueron el industrial, materiales, servicios financieros y telecomunicaciones. El riesgo de todos los índices sectoriales durante la crisis de salud de 2020 ha sido de menor cuantía en comparación con el riesgo que tuvieron durante las crisis de 2008. En lo general, los resultados de esta investigación pueden ser útiles para la diversificación de portafolios de inversión y, de manera más general, para diseñar estrategias de administración de riesgos de esas inversiones. De manera particular sugieren que las crisis mundiales no siempre tienen efectos iguales y que no sólo su origen, sino su naturaleza específica debe formar parte de la agenda de investigación.