These findings support selecting forecasting models according to operating conditions rather than relying on a single universally preferred approach, and provide a practical framework for combining complementary statistical and machine-learning forecasts.
Abstract
Public health forecasts must respond to abrupt changes in surveillance data without over-extrapolating noise, reporting artifacts, or temporary trends. We evaluated autoregressive integrated moving average (ARIMA), random forest, and extreme gradient boosting (XGBoost) models using 190 weekly observations of publicly available Ontario COVID-19 case counts from January 2020 to October 2023. Rolling-origin time-series cross-validation preserved temporal order during model tuning and evaluation. Performance was assessed across three operating dimensions: responsiveness following selected turning points, forecast horizons of one to six weeks, and the amount of historical training data. We also developed Machine Learning and ARIMA Model Averaging (MLAMA), a non-negative performance-weighted ensemble with weights that vary by forecast horizon and responsiveness setting. Retrospective comparisons showed that ARIMA adapted rapidly after turning points but its normalized error increased at longer horizons. Random forest and XGBoost were less responsive initially but maintained more stable normalized error over longer horizons. For two-week forecasts at the end of the study period, training on the most recent data outperformed using longer historical periods, particularly for XGBoost. MLAMA achieved the lowest normalized mean absolute percentage error across most forecast horizons and ranked among the best-performing methods across responsiveness settings. These findings support selecting forecasting models according to operating conditions rather than relying on a single universally preferred approach. MLAMA provides a practical framework for combining complementary statistical and machine-learning forecasts. The accompanying Python package is currently maintained in a private repository while software validation and reproducibility testing are completed.
AbstractConventional Vector Autoregressive (VAR) models are widely applied for multivariate time series analysis, their performance deteriorates in high-dimensional settings due to inefficient parameter estimation, unstable forecasts, and difficulties in interpreting temporal dependencies. This study conducts a comparative study on conventional Vector Autoregressive (VAR) models, Multivariate Random Forest for VAR (MRF-VAR) models and Multivariate Extreme Gradient Boosting for VAR (MXGB-VAR) models, validated using simulated and real-life dataset for Nigerian financial time series. Augmented Dickey–Fuller (ADF) test was adopted to test the stationarity of the data. Forecast accuracy across short-term and long-term horizons for the models were measured using Mean Absolute Deviation (MAD) and Root Mean Square Deviation (RMSD). Results for simulated data show that the conventional VAR models achieved the best short-term forecast performance (MAD = 1.642, RMSD = 2.016), while the MRF-VAR models performed best in long-term forecasting (MAD = 0.947, RMSD = 1.197). In the case of the real-life dataset for Nigerian financial time series, the MRF-VAR model performed well in short-term forecasts (MAD = 108.84, RMSD = 149.53), whereas MXGB-VAR model provided better results in long-term forecasts (RMSD = 730.57). Policymakers and financial analysts should be encouraged to apply machine learning approaches to VAR models in macroeconomic and financial forecasting to improve decision-making.
N. Isah, S. Doguwa· Indian Journal of Industrial...· 0 citations
SARIMA’s versatility in capturing weekly seasonal patterns, as observed in the 7-day cycles within the COVID-19 data, which reflect reporting or behavioral trends, is highlighted.
M. Sorokina, I. Korshukov, N. Omarbekova et al.· Medicine and ecology· 0 citations
The Hybrid ARIMA–LSTM model consistently achieved superior forecasting accuracy, with lower errors and better goodness-of-fit than the standalone models, while supporting the Hybrid ARIMA–LSTM framework as a robust tool for long-term epidemiological forecasting.
D. Singh· International Journal For Mu...· 0 citations
Machine learning models that predict hospital admission at triage may support patient flow forecasting, yet the effects of covariate drift, concept drift, and retraining on long-term performance are poorly understood. We developed an Extreme Gradient Boosting (XGBoost) model using deidentified data from all presentations to a metropolitan hospital in Western Australia. Training and validation included 2016 and 2017 presentations (n = 133,814), with rolling quarterly testing from 2018 to 2023 (n = 455,496). Two adaptive strategies were evaluated: quarterly and half-yearly retraining. Covariate drift was assessed using univariate and multivariable analyses, and reporting adhered to TRIPOD + AI and MINIMAR standards. Substantial drift was observed both between training and testing datasets and across the six-year testing period. The base model achieved a mean AUROC of 0.875 (range 0.844-0.887) and mean daily bed error of 7.42 beds (range 0.47-13.1). Retrained models demonstrated improved discrimination (mean AUROC 0.892 and 0.893) and reduced bed error (4.02 and 4.48 beds per day) for quarterly and half-yearly retraining, respectively, with similar calibration and classification performance. Covariate drift meaningfully degraded calibration but not discrimination metrics over time. Simple retraining improved discrimination and reduced calibration concept drift, underscoring the importance of retraining to address drift for temporal model deployment.
Ethan Williams, Toshi Sinha, M. Summerscales et al.· npj Digital Medicine· 0 citations
ML inflation forecasts are almost universally trained on fully revised data, even though real-time forecasters never have such data, and reported feature importances are typically computed in-sample, conflating predictive relevance with retrospective fit. This paper asks whether the ML advantage over the Phillips curve documented in Agyekum (2026) survives when models are trained and evaluated on real-time (ALFRED) vintages rather than revised series, and whether SHAP feature-importance rankings are an artifact of in-sample estimation. Using 2000-2026 U.S. data on unemployment, CPI and PCE inflation, payrolls, real GDP, and the 10-year-2-year Treasury spread, vintage-consistent panels are built for four traditional models (random walk, AR(1), Phillips curve, ADL-OLS) and four ML models (Random Forest, Gradient Boosting, Elastic Net, SVR), re-estimated recursively at 3-, 6-, and 12-month horizons (208, 206, 204 forecasts). Real-time/revised accuracy differences are small and, apart from one exception at 6 months (Gradient Boosting vs. Phillips curve, DM = -1.671, p = 0.097), indistinguishable under Diebold-Mariano tests; Gradient Boosting alone shows consistent positive skill at longer horizons. The random walk remains a strong short-horizon benchmark, consistent with the puzzle in Agyekum et al. (2026) for exchange rates. Using walk-forward, out-of-sample SHAP, a Random Forest on revised data assigns dominant importance to PCE inflation (mean |SHAP| = 0.778, rank 1 of 9), while on real-time data it assigns PCE negligible importance (0.039, rank 6), relying instead on current CPI (0.834 vs. 0.223). This twenty-fold swing, larger than the in-sample estimate, is invisible to point-forecast metrics and shows the model's PCE reliance is substantially a hindsight artifact. An RSI summarizes the accuracy gap by model and horizon, with implications for auditing ML inflation forecasts.
A weeklong summer workshop brought higher education faculty to campus to explore how AI and machine learning materials can be adapted for their classrooms.
