Antonio Pagliaro has a total of 1000 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Can Scene-Level NWP Cloud Cover Replace Expert Pixel Labels? A Large-Scale Assessment with ERA5, GFS and CloudSEN12+

    by Anna Anzalone, Antonio Pagliaro

    Deep learning cloud detectors require large volumes of pixel-level labels that are expensive to produce. Gridded cloud cover from ERA5 reanalysis and GFS forecasts could serve as freely available weak supervision. We assess this hypothesis through a large-scale comparison of NWP total cloud cover with expert-annotated cloud masks from CloudSEN12+ (Sentinel-2, 10m) over 48,000 globally distributed samples (2017–2022). ERA5 shows moderate agreement (r=0.42, MAE ∼30%), with performance strongly dependent on cloud type. We then train U-Net cloud segmentation models that differ only in the source of their labels, on a region-disjoint partition of the dataset, and evaluate them on 975 held-out expert-annotated patches. Expert labels yield a binary cloud IoU of 0.78 and a scene-level accuracy of 0.92, against 0.47 and 0.71 for ERA5-derived weak labels. Increasing the weak supervision from 200 to 20,000 patches produces no systematic improvement over the tested range, which suggests that, for this label formulation, the limitation lies in what each label expresses rather than in how many are available; 849 expert-annotated patches outperform 20,000 weakly labelled ones. Pretraining on ERA5 labels and fine-tuning on expert masks is worse than training from scratch at every annotation budget that we tested. Training on ERA5 labels does improve on applying the same labels directly as a mask, from 0.38 to 0.47 in binary cloud IoU, and the resulting model remains usable for coarse scene-level screening where no pixel-level annotation exists. Results provide practical guidelines for leveraging NWP products in deep learning training pipelines where expert annotations are scarce.

  • Open Access

    When Feature Alignment Helps, Hurts, or Washes out: Carrier-Dependent Domain Adaptation for Cherenkov Telescope Event Classification

    by Antonio Pagliaro

    Feature alignment domain adaptation (DA) methods such as CORAL and MMD are widely used to mitigate distribution shift between training (source) and deployment (target) domains. They are typically evaluated on a single downstream classifier, which implicitly assumes that the alignment benefit transfers across model families. We test this assumption on a controlled class-symmetric shift of Imaging Atmospheric Cherenkov Telescope features, with five carrier classifiers, three alignment methods, and fifteen seeds per cell. Establishing this requires first correcting an evaluation protocol that we believe to be common in benchmarks built by perturbing a single sample: when the target consist of a perturbed copy of data that the carrier was partly trained on, the measured benefit of adaptation is inflated by a factor that tracks the carrier’s capacity to memorise (from 1.0 for logistic regression to 5.7 for an unbounded-depth random forest), so that such a benchmark reports a carrier gradient whether or not one actually exists. When scored only on events that no carrier has seen, the carrier still matters (indeed, twice over): the same shift costs a hyperplane carrier 0.004 of AUC and a tree ensemble 0.006–0.008, and CORAL then recovers 58% of that loss on logistic regression, against 84–98% on the tree ensembles, a difference that RBF-kernel MMD does not show on any carrier for which the ratio is determined. A measurement of marginal sensitivity taken without any adaptation reproduces the ordering. Two effects that would be reported by a full-target protocol are absent on held-out events: a benefit growing with tree depth, and an advantage of covariance alignment over mean matching on deep trees. We propose a two-axis framework (shift symmetry × carrier marginal sensitivity), then test its class-symmetric row and give practical recommendations for evaluating and deploying domain adaptation under shift.

  • Open Access

    A quantitative benchmark for explainable artificial intelligence in astronomical transient classification

    by A. Pagliaro

    Explainable artificial intelligence (xAI) methods are increasingly used to interpret machine learning classifiers of astronomical transients, yet their reliability is rarely assessed quantitatively. We present astro-xai-bench , an open benchmark that evaluates explanation quality along five complementary axes – perturbation-based faithfulness, deletion and insertion AUC, complexity, cross-representation plausibility, and the standard model- and data-randomisation sanity checks – applied to nine classifiers spanning tree-based, deep-learning, and physics-informed architectures on the PLAsTiCC photometric transient dataset (14 classes, six passbands), with all metrics averaged over five random seeds. A detailed case study of XGBoost (tree-based) and CNN2D (convolutional on Gramian angular field images) illustrates how each xAI axis interacts with model architecture: faithfulness is representation-dependent, complexity rankings invert under basis change, and projected plausibility is sensitive to the projection rather than the model. Three general findings emerge: (i) deep-learning models on raw light curves align better with an astrophysically informed reference mask than tree-based models trained directly on the hand-crafted features that define the mask; (ii) the sanity checks reveal that LSTM, CNN1D, and Random Forest attributions are dominated by the input distribution, while XGBoost, CNN2D, ViT, and both physics-informed baselines (Avocado, ParSNIP) pass all applicable checks; (iii) no single model dominates all five axes, making multi-axis evaluation essential. All code and evaluation pipelines are released as a public benchmark.

  • Open Access

    The ASTRI-Horn Cherenkov Telescope: Technical Assessment and Calibrations Following its Refurbishment for the ASTRI Mini-Array and CTA-SST Pathfinder

    by G. Leto, C. Bigongiari, M. Capalbi, O. Catalano, V. Conforti, G. Contino, A. Compagnino, S. Crestan, F. Gianotti, A. Giuliani, et al.

    The ASTRI-Horn telescope is a prototype Imaging Atmospheric Cherenkov Telescope (IACT) developed within the Astrofisica con Specchi a Tecnologia Replicante Italiana (ASTRI) project. It represents the first implementation of a dual-mirror Schwarzschild–Couder (SC) optical configuration for ground-based gamma-ray IACT observations. It also served as a testbed for the Small-Sized Telescopes (SSTs) of the Cherenkov Telescope Array Observatory (CTAO) that will host at least 37 SSTs at the Paranal site in Chile, and for the forthcoming ASTRI Mini-Array. Installed in 2014 at the INAF “M.C. Fracastoro” observing station on Mount Etna, the telescope is equipped with a SiPM-based Cherenkov camera and a complete control and data acquisition system. Between 2020 and 2022, after the demonstration of the validity of the SC configuration in observing gamma-ray sources, ASTRI-Horn underwent refurbishment activities that included mirror replacement, new optical coatings, camera upgrades, and software improvements. The telescope has since been re-validated through calibration campaigns and observational tests, including the detection and spectral characterization of the Crab Nebula at TeV energies. The results confirm the correct operation of the optical system, camera electronics, and data processing chain, providing a consolidated basis for the design and development of the ASTRI Mini-Array and future CTAO-SST telescopes.

  • Open Access

    Chasing Gamma-Ray Signals from Binary Neutron Star Coalescences with the Cherenkov Telescope Array: Prospects and Observing Strategies

    by S. Abe, J. Abhir, A. Abhishek, F. Acero, A. Acharyya, R. Adam, A. Aguasca-Cabot, I. Agudo, I. Albanese, J. Alfaro, et al.

    The detection of gravitational waves (GWs) from a binary neutron star (BNS) merger by Advanced LIGO and Advanced Virgo (GW170817), together with its electromagnetic counterpart, the short gamma-ray burst GRB 170817A, heralded the birth of multimessenger astronomy. The detection of TeV emission from GRBs motivates follow-up observations with the Cherenkov Telescope Array Observatory (CTAO), which is ideal for detecting such signals due to its unprecedented sensitivity, rapid response, and wide-field survey capabilities. The aim of this work is to evaluate GeV–TeV GW follow-up strategies for CTAO using a multistep simulation pipeline and to estimate the expected rate of joint GW–GRB detections during observing run O5. Using a simulated sample of BNS systems with corresponding GW detections, gamma-ray emission is simulated through phenomenological prescriptions based on the observed population of short GRBs, including off-axis jet scenarios. CTAO observations are simulated to account for instrument response, sky tiling strategies, integration times, and varying observing conditions. Strategies with variable and constant integration times are investigated. We find that, via an optimized follow-up strategy, about 5% of simulated GW-associated short GRBs produce GeV–TeV radiation detectable by CTAO. Detectability is strongly influenced by the jet opening angle and viewing angle, suggesting that even rough estimates of the viewing angle in GW alerts could enhance targeting. This framework motivates future follow-ups of GW-detectable events, including neutron star–black hole mergers, and further supports the development of advanced strategies incorporating galaxy distributions and synergies with future detectors such as the Einstein Telescope.

  • Open Access

    Cognitive Biases in Large Language Models: A Systematic Quantitative Assessment and Debiasing Analysis

    by Antonio Pagliaro

    Large Language Models (LLMs) are increasingly deployed in decision-support systems across high-stakes domains, yet their susceptibility to cognitive biases—systematic deviations from rational judgment well-documented in human psychology—remains poorly understood in quantitative terms. Existing studies typically examine a narrow set of biases, test a single model family, and rely on qualitative assessments of bias presence. In this work, we present a rigorous experimental framework, inspired by the methodology of experimental physics, for the systematic quantitative measurement of cognitive biases in LLMs. We introduce the Bias Strength Index (BSI), a normalized metric with associated confidence intervals that quantifies the magnitude of bias on a continuous scale, and we decompose the total uncertainty into statistical and systematic components—the latter arising from prompt reformulation. We evaluate a comprehensive taxonomy of eleven cognitive biases (including anchoring, framing effect, confirmation bias, availability heuristic, sunk cost fallacy, bandwagon effect, status quo bias, and others) across eight state-of-the-art LLMs from seven families: GPT-4.1 Mini, Claude 3.5 Sonnet, Gemini 2.5 Flash, Llama 3.3 70B, Llama 3.1 8B, Mistral Large (mistral-large-2411), DeepSeek V3, and MiniMax M2.5. Each bias is probed through multiple semantically equivalent prompt variants, with N = 100 independent trials per configuration, yielding a dataset of over 70,000 model responses. Our results reveal that all tested models exhibit non-zero bias effects for multiple bias categories, though with markedly different profiles. A trial-level Generalized Linear Mixed-Effects Model (GLMM) analysis finds statistically significant bias effects in 27 of 43 testable bias–model combinations (62.8%) after multiple-comparison correction, while a more conservative variant-level test—which requires effects to generalize across prompt formulations—yields only one significant result, highlighting the dominant role of prompt-induced systematic uncertainty. Framing and primacy/recency effects are near-universal, while susceptibility to other biases varies substantially across model families. We further evaluate three debiasing strategies—zero-shot chain-of-thought, adversarial counter-prompting, and role-based prompting—applied at inference time without modifying model weights. Our findings provide a quantitative foundation for auditing cognitive biases in LLMs and highlight the bias-dependent effectiveness of prompt-based debiasing techniques.


  • Top Cited Papers
    Open Access

    Regime-Aware LightGBM for Stock Market Forecasting: A Validated Walk-Forward Framework with Statistical Rigor and Explainable AI Analysis

    by Antonio Pagliaro

    Can machine learning generate statistically validated alpha in equity markets while adapting to changing market conditions? This study addresses this question by proposing a regime-aware LightGBM framework conditioned on market regimes detected via a rolling Hidden Markov Model, eliminating look-ahead bias. Backtested on 51 NASDAQ-100 constituents (2015–2026), the strategy achieved a portfolio Sharpe ratio of 1.18 (95% CI: [0.53, 1.84]) and outperformed four baseline models. The key findings include the following: (i) cross-asset features (Bitcoin as a leading indicator) contribute the most predictive value; (ii) macroeconomic indicators outweigh traditional technical indicators for high-beta stocks; (iii) the model autonomously adapts its decision logic across regimes, shifting from mean reversion in bear markets to risk appetite monitoring in bull markets. While block bootstrap tests confirm statistical significance (p<0.001), the Deflated Sharpe Ratio (0.69) does not reach formal significance after multiple testing correction—an honest finding we report transparently.


  • Open Access

    Cognitive Biases in Asset Pricing: An Empirical Analysis of the Alphabet Effect and Ticker Fluency in the US Market

    by Antonio Pagliaro

    Behavioral finance theory predicts that Processing Fluency—the subjective ease of parsing a nominal stimulus—should systematically influence investor attention and asset pricing through heuristic-based decision making. Yet modern equity markets, increasingly dominated by High-Frequency Trading (HFT) and algorithmic execution, provide powerful near-instantaneous arbitrage forces that should neutralize any pricing premium arising from superficial nominal cues. Whether cognitive biases such as the “Ticker Fluency” effect and the “Alphabet Effect” persist in this algorithmic environment or have been fully arbitraged away remains an open empirical question with direct implications for the boundary conditions of Processing Fluency Theory. We address this gap by applying a deterministic Heuristic Fluency Score—based on vowel density and consonant cluster penalties—to all 492 S&P 500 constituents over 752 trading days (January 2021–January 2024), estimating individual stock Fama-French 3-Factor Alphas via daily time-series regressions, and testing whether fluency or alphabetical rank explains cross-sectional variation in abnormal returns after controlling for Liquidity, Amihud illiquidity, and GICS Sector Fixed Effects. To guard against Selection Bias, we explicitly contrast a biased illustrative case study (N=25, 2019–2024) against the rigorous full-market analysis. We find no statistically or economically significant effect: the Fluency Score coefficient is β=0.0036 (p=0.495) and the Alphabet Rank coefficient is β=−0.0027 (p=0.642), with the results robust to all tested parameterizations (λ∈[0.05,0.20]; p>0.50 throughout). These findings establish a boundary condition of Processing Fluency Theory: in algorithm-dominated, highly liquid large-cap markets, cognitive biases in nominal cues are fully absorbed by arbitrage, and ticker symbols function as neutral identifiers rather than heuristic signals. Residual effects, if any, are more likely to manifest in attention-based or volume-related outcomes, or in less institutionalized market segments where algorithmic participation is lower.


  • Open Access

    The Specialization of Intelligence in AI Horizons: Present Status and Visions for the Next Era

    by Antonio Pagliaro, Pierluca Sangiorgi

    The current era of artificial intelligence is often portrayed through the lens of large general-purpose models

  • Open Access

    An Introduction to Machine Learning Methods for Fraud Detection

    by Antonio Alessio Compagnino, Ylenia Maruccia, Stefano Cavuoti, Giuseppe Riccio, Antonio Tutone, Riccardo Crupi, Antonio Pagliaro

    Financial fraud represents a critical global challenge with substantial economic and social consequences. This comprehensive review synthesizes the current knowledge on machine learning approaches for financial fraud detection, examining their effectiveness across diverse fraud scenarios. We analyze various fraud types, including credit card fraud, financial statement fraud, insurance fraud, and money laundering, along with their specific detection challenges. The review outlines supervised, unsupervised, and hybrid learning approaches, discussing their applications and performance in different fraud detection contexts. We examine commonly used datasets in fraud detection research and evaluate performance metrics for assessing these systems. The review is further grounded by two case studies applying supervised models to real-world banking data, illustrating the practical challenges of implementing fraud detection systems in operational environments. Through our analysis of the recent literature, we identify persistent challenges, including data imbalance, concept drift, and privacy concerns, while highlighting the emerging trends in deep learning and ensemble methods. This review provides valuable insights for researchers, financial institutions, and practitioners working to develop more effective, adaptive, and interpretable fraud detection systems capable of operating within real-world financial environments.