This study forecasts unusually large daily Bitcoin price movements and examines whether the resulting warnings improve investment decisions. We compare simple volatility signals, a benchmark based on the previous day’s return magnitude, logistic regression, and random forests in chronological tests covering 2340 days. Large movements include gains and losses whose absolute daily log returns exceed the 95th percentile of the preceding training sample. The evaluation contains 51 such events. The selected volatility signal issues 77 warnings and captures nine events. Better predictive ranking does not consistently produce better warnings. Additional comparisons use model tuning, common nominal warning budgets, and a separate analysis of large losses. Investment evaluation accounts for trading costs, cash returns, execution delays, and investor preferences. Moving fully to cash after the selected loss warnings can reduce investor utility relative to buying and holding Bitcoin. In contrast, adjusting exposure using volatility over the preceding 20 days improves utility at the highest tested risk aversion, despite lowering total return. These historical comparisons are exploratory; uncertainty estimates condition on recorded forecasts and are not adjusted for multiple comparisons. The findings support separate evaluation of predictive ranking, warning governance, and economic value.
Forecasting food costs is an important tool for food security planning and management. This study aimed to evaluate and compare the performance of statistical and machine learning models for forecasting the monthly cost of the Basic Family Food Basket (BFFB) in Panama and San Miguelito. Monthly data from 2009 to 2024 were chronologically divided into training and test sets. A Naïve model was used as a benchmark, while ARIMA, Holt, Random Forest and XGBoost models were evaluated using MAE, RMSE and MAPE. The main evaluation used a fixed-origin multi-step forecasting scheme, complemented by a rolling-origin evaluation. Predictive accuracy was also assessed using the Diebold–Mariano test with Holm correction for multiple comparisons. Random Forest yielded the lowest test-set errors (MAE = 13.82, RMSE = 17.23, MAPE = 3.92%), followed by XGBoost (MAE = 15.84, RMSE = 18.14, MAPE = 4.53%). The rolling-origin evaluation also showed favorable performance for both machine learning models, while the Diebold–Mariano test identified statistically significant differences in four of the ten pairwise comparisons based on the Holm p-values, all involving XGBoost. These results highlight the potential of machine learning models for BFFB cost forecasting and support their usefulness as tools for food security-related monitoring and planning.
Cross-group comparisons require both adequate measurement-model estimation and a defensible procedure for expressing group-specific estimates on a common metric. This study presents a reproducible workflow for documenting the operational feasibility of multi-group linking in short ordinal scales. Data were collected from 373 students enrolled in Portuguese higher education who completed a 21-item entrepreneurial behaviour questionnaire comprising seven three-item dimensions. For each dimension and each of three grouping variables—gender, field of study and nationality—the workflow applied subgroup-size screening, response-category variation checks, common candidate-item screening, group-specific graded response model estimation, Stocking–Lord linking and stage-specific diagnostic reporting. All 21 group-by-dimension configurations satisfied the implemented screening rules for subgroup support, item variation and common candidate-item availability. However, operationally linked estimates were obtained for only 16 configurations. Four configurations stopped because at least one required group-specific graded response model did not converge, and one stopped because the Stocking–Lord optimisation failed. Sensitivity analyses indicated that convergent linking solutions were highly stable across numerical starting values but could change substantially when one item was removed from the three-item linking set. The resulting linked parameters and latent scores should therefore be interpreted as conditional on the fitted models, assumed common-item sets, selected reference groups and implemented estimation procedures, rather than as evidence of measurement invariance or absence of differential item functioning. The study’s principal contribution is a transparent analytical framework that distinguishes initial screening feasibility, group-specific model-estimation success and operational metric alignment, and treats subsequent cross-group contrasts solely as diagnostic illustrations of the workflow rather than as substantive findings about gender, field of study or nationality.
This study develops region-specific threshold-based class definitions for recorded forest fire incidents in Attica, Eastern Macedonia and Thrace, and Crete, Greece, using meteorological and operational data from 2000–2024. Regional temperature, relative humidity, and wind-speed thresholds were compared with the conventional 30/30/30 rule, requiring all three criteria to be satisfied simultaneously. The regional criteria were satisfied in 46.2%, 49.8%, and 35.5% of long-duration incidents (>360 min) and 22.4%, 26.9%, and 24.8% of short-duration incidents, respectively. Corresponding proportions under the conventional rule were 5.2%, 0%, and 0.3% for long-duration incidents and 1.5%, 0%, and 0.4% for short-duration incidents. Because thresholds were derived from the same datasets, this comparison is in-sample and exploratory. Six machine-learning approaches were compared for reproducing the region-specific threshold-based classes: MLP, RBF, Random Forest, FC2 (RF), FC2 (MLP) and GENCLASS. The FC2 variants use Grammatical Evolution for feature construction, whereas GENCLASS performs evolutionary rule induction. GENCLASS achieved the lowest average classification error overall, followed by FC2 (MLP). However, with only three regional datasets, the Friedman test indicates overall differences, but only the difference between GENCLASS and RBF is statistically significant after correction for multiple comparisons. These classification results concern reproduction of the defined classes, not independent prediction of fire duration. Independent incident data are required to validate duration-related findings; evaluating fire occurrence additionally requires fire-free observations.
Background: Can predictive algorithms trained on survey data affected by systematic measurement error meet official statistics quality standards? This study examines the tension between design-based inference and algorithmic modelling in poverty estimation, focusing on regressive income underreporting in the Italian shadow economy. Methods: A Monte Carlo microsimulation protocol of 1000 replications is calibrated on ISTAT and EU-SILC 2025 statistics. The shadow economy is formalised as Missing Not At Random measurement error with spatial heteroscedasticity across three Italian macro-regions. Five estimators—design-based, algorithmic, and hybrid—are compared and audited through the QOSU framework. Results: The GREG retains asymptotic design-unbiasedness on undistorted data but exhibits bias up to +29% on observed data, showing that calibration is insufficient under non-ignorable error. Uncalibrated machine learning estimators display systematic bias via shrinkage or amplification, termed statistical hallucination. The Asymmetric Hybrid, with a parameter-free quadratic weighting function, achieves GOLD in five of six domain–indicator combinations; a hold-out validation identifies a reliability threshold of 0.70 as the framework’s domain of validity. Conclusions: Integrating machine learning into official statistics requires an adaptive architecture in which each component’s contribution is modulated by the reliability of the information base. The QOSU framework provides a reproducible protocol for auditing algorithmic estimates under non-ignorable measurement error.
