Ignacio Crespo

This paper introduces the Fluctuant-SPA (FSPA) test, extending the Superior Predictive Ability (SPA) test to unstable environments. The FSPA asks whether a benchmark outperforms all alternatives in every pre-specified sub-period jointly, rather than on average, making it sensitive to time-varying relative predictive ability. We apply the test to downside risk forecasts for the U.S. economy over 1984–2025, comparing different competing methodologies against a financial-conditions quantile regression benchmark. Performance rankings shift substantially across sub-periods, with methods dominating in some episodes but underperforming in others. At the annual horizon the evidence is borderline: the FSPA rejects the benchmark's dominance while the SPA, averaging over the full sample, does not. The divergence is sharpest when the evaluation ends before the COVID-19 pandemic. At the one-quarter horizon, the benchmark's dominance survives both tests, but the FSPA trajectory shows it emerged only with the Global Financial Crisis, a pattern invisible to the global SPA.

Earlier versions were presented at 2023 BSE Summer Forum (Workshop on Macroeconomics and Policy Evaluation; poster session), 3rd International Econometrics PhD Conference (Erasmus University Rotterdam), 44th International Symposium on Forecasting (Dijon), and SoFiE Summer School “Monitoring and Forecasting Macroeconomic and Financial Risk” (Brussels).

Intra-daily trading volume forecasts are a key input for algorithmic trade execution. We introduce a forecasting methodology for large panels of assets that combines factor models with sparse vector autoregressions. The approach captures both market-wide common factors driving trading activity and the sparse network of volume spillovers among individual assets, consistent with a framework where information diffuses sequentially from leader to lagger stocks. We apply the methodology to a panel of 600 constituents of the STOXX 600 index spanning 16 European countries. We assess both the statistical accuracy and the economic value of the forecasts through an out-of-sample prediction exercise and a VWAP trade execution analysis. Results show that our methodology delivers significant statistical and economic gains relative to standard univariate benchmarks. The improvements are heterogeneous: the largest gains accrue to stocks with high network connectivity, while firm size also plays a significant role, consistent with the information diffusion mechanism underlying our approach.

We introduce a test for whether a benchmark method remains superior uniformly across competing alternatives for a panel of target series, and apply it to macroeconomic forecasting using the FRED-MD database. We find that the factor-model benchmark proves difficult to outperform uniformly when competing against alternatives spanning univariate methods, penalized regressions, and machine learning. Our findings point to the value of uniform testing when assessing predictive ability across series.

Presented at 29th Finance Forum (AEFIN), 5th Annual Workshop on Financial Econometrics (Örebro), 27th Meeting of Young Economists (SMYE 2023), and 47th Meeting of the Brazilian Econometric Society (Insper, São Paulo).