The study, conducted by Damian Machlanski of the University of Edinburgh, Spyridon Samothrakis of the University of Essex, and Paul Clarke of the University of Essex, tackles a puzzle at the heart of what statisticians call treatment effect estimation. Using Rubin’s potential outcomes framework, the goal is to estimate the difference between what happened to people who received a treatment and what would have happened to those same people had they not received it. The average of these differences across a population is the Average Treatment Effect, or ATE, while the conditional version, CATE, describes how the effect varies across subgroups. The catch is that one of the two potential outcomes for every individual is always counterfactual—it never happened and can never be observed. Unlike predicting house prices or classifying images, where a validation set reveals exactly how wrong a model is, causal models are graded on answers that are fundamentally hidden.

This hiding is what makes hyperparameter tuning so treacherous in the causal setting. In standard supervised machine learning, everyone agrees on the metric: minimize mean squared error on held-out data. The best-fitting model wins. In causal inference, what researchers actually care about is the error on the difference between potential outcomes, quantified by metrics like the Precision in Estimation of Heterogeneous Effects (PEHE). But because counterfactuals are unobservable, only the observable Mean Squared Error on factual outcomes is available in real datasets. The team formalizes this tension by distinguishing the observable metric, which can be computed, from the potential metric, which is ideal but inaccessible. When treatment assignment is random, as in a randomized controlled trial, the observable error is a decent stand-in for the potential one. Observational data, however, is riddled with covariate shift—the characteristics of treated and untreated individuals differ systematically—and the proxy can fail spectacularly.

To demonstrate the problem vividly, the authors construct a deceptively simple toy example: a one-dimensional input, a sinusoidal data-generating process, and a treated group with missing regions of data. Three candidate models emerge. The first fits the control arm beautifully but misses the trend in the treated arm, achieving the best observable MSE while producing terrible causal estimates. The second captures the true trend in both arms despite the data gaps—the solution anyone would want—but its MSE is worse, so conventional model selection would reject it. The third fits the data poorly overall yet, through leakage of benchmark test information, manages to top the PEHE leaderboard. All three illustrate the same disturbing conclusion: in causal estimation, the metric used to pick a model can be actively misleading, and a solution that looks best on paper may be the wrong one.

The centerpiece of the paper is an exhaustive empirical study built on the four classic causal benchmarks: IHDP, derived from the Infant Health and Development Program; Jobs, combining experimental and observational employment data; Twins, based on twin births in the United States between 1989 and 1991; and News, a high-dimensional dataset of simulated user responses to articles. Crucially, these benchmarks contain ground-truth counterfactuals, which allowed the researchers to compute the otherwise inaccessible potential metrics—not to tune models, which would violate good practice, but to grade how well different observable metrics perform as tuning proxies. The experimental grid spans seven widely used causal estimators, including S-Learners, T-Learners, X-Learners, Doubly Robust estimators, Double Machine Learning, Inverse Propensity Score Weighting, and Causal Forests, crossed with nine base learners ranging from Lasso and Ridge regression to gradient-boosted trees and neural networks.

The headline result is stark. Using an oracle that selects hyperparameters with the ideal potential metrics—the best any data-driven tuning could ever achieve—the probability of reaching state-of-the-art performance jumped from 65 percent to 81 percent for average treatment effect estimation, and from 50 percent to 57 percent for individualized estimation. The comparison held across nearly every estimator and base learner, with typical gains of 8 to 12 percent, and even larger improvements of 10 to 33 percent for learners like Lasso, Decision Trees, Kernel Ridge, and CatBoost. Only the ensemble methods, including Random Forests and Causal Forests, showed robust defaults—a likely explanation for the strong out-of-the-box performance of Causal Forests, whose ensembling structure echoes the resilience of the other ensemble learners.

The implications ripple outward. If swapping hyperparameters alone can push an estimator from the bottom of the league table to state-of-the-art territory, then published benchmarks that compare causal methods may be measuring tuning quality as much as algorithmic merit. The authors point out that different studies use different metrics for hyperparameter optimization, making cross-paper comparisons of estimators inherently fragile. Their conclusion is deliberately provocative: a practitioner who tunes a single model carefully may outperform a rival who searches over many models without tuning. In other words, the casual assumption that newer, more sophisticated neural estimators inherently beat older regression-based approaches may simply reflect differences in how diligently their knobs were turned.

Yet the study also offers a measure of hope. Not every observable metric fails equally. When the team evaluated several practical tuning proxies—including factual MSE, plugin validation that synthesizes pseudo-ground-truth effects with a second estimator, k-nearest-neighbour matching, and the R-Score derived from the R-Learner framework—they found that some came remarkably close to oracle performance. Matching-based plug-in metrics in particular showed low variability, with a 95 percent confidence interval of roughly 0.01 on the aggregated probability scale, suggesting that well-chosen proxies can put excellent hyperparameters within reach. Others fared badly: the R-Score behaved erratically, diverging from recent findings, and plain MSE, while on average close to the oracle, produced inconsistent results across scenarios. The choice of metric, the authors conclude, can seriously distort final estimation performance, and there is currently no metric uniformly capable of identifying the best causal model.

The researchers are candid about the limits of their work. The analysis rests on the standard assumptions of causal inference—no interference between units, no hidden confounders, and positive probability of treatment for everyone—and the search spaces, while broad, had to be trimmed for computational feasibility. Still, their practical recommendation is strikingly modest: until more robust observable metrics are developed, practitioners may be better served by highly interpretable models such as linear regression or decision trees, whose hyperparameters domain experts can tune manually and thereby embed genuine prior knowledge. Perhaps most provocatively, they argue that the field does not need new estimators at all. Meaningful progress, they contend, will come from a deeper understanding of causal model selection and hyperparameter tuning—the same realization that transformed general machine learning a decade ago—potentially aided by standardized public benchmarking platforms that tune every causal estimator equally thoroughly, so that claimed breakthroughs can finally be credited to the methods themselves rather than to lucky settings.

Subject of Research: Hyperparameter tuning and model selection metrics for machine learning-based causal effect estimation

Article Title: The Challenges of Hyperparameter Tuning for Accurate Causal Effect Estimation

Article References: Machlanski, D., Samothrakis, S., & Clarke, P. (2026). The Challenges of Hyperparameter Tuning for Accurate Causal Effect Estimation. Machine Learning, 115(11), Article 249. https://doi.org/10.1007/s10994-026-07140-9

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07140-9

Keywords: causal inference, hyperparameter tuning, treatment effect estimation, CATE, model selection, machine learning, benchmarks, evaluation metrics, double machine learning, observational data, counterfactuals, Challenges

The team’s starting point was a rich dataset of 72 real-driving trips performed with a BMW i3, a compact electric vehicle whose battery and heating circuit data were recorded under genuine operational conditions. This kind of data is gold for battery researchers. Synthetic drive cycles, such as the standardized profiles used in certification tests, smooth away the stop-and-go chaos, temperature swings, and driver behavior that dominate everyday motoring. Previous comparative studies have shown that battery degradation patterns on synthetic cycles can differ markedly from those observed in the wild, which means control strategies tuned in the laboratory often underperform on the street. By anchoring their models to 72 authentic journeys, Chen and Zhou ensured that every prediction and optimization they made reflected the messy reality of daily driving.

The first technical step was validation. The researchers constructed detailed vehicle and heating circuit models and checked them against the recorded trip data, confirming that their simulations reproduced the thermal and electrical behavior of the actual car. With trustworthy models in hand, they then developed strategies for a clever energy trick: using regenerative braking energy to heat the cabin. In an electric vehicle, braking normally wastes kinetic energy as heat at the friction pads, but regenerative systems recover part of it as electricity. Routing that recovered energy into cabin heating rather than drawing it from the traction battery eases the load on the cells, and the study reports that such strategies improved both the vehicle’s driving range and battery life. It is a double win: passengers stay warm, and the battery endures fewer stressful charge-discharge excursions.

From there, the study pivoted to machine learning. The researchers framed a prediction problem in which environmental variables, vehicle variables, battery variables, and thermal system variables serve as inputs, while the target outputs are the battery’s final State of Charge and the distance traveled in kilometers. In plain terms, they wanted an algorithm that could look at the conditions of a trip and accurately forecast how much energy would remain in the battery and how far the car would go. Four candidate models were systematically compared: XGBoost, AdaBoost, CatBoost, and K-Nearest Neighbor. Each was assessed with a rigorous protocol, an 80/20 train-test split combined with five-fold cross-validation, which guards against the common pitfall of an algorithm that merely memorizes its training data.

Among the four contenders, CatBoost, a gradient-boosting method developed originally to handle categorical features gracefully, emerged as the strongest baseline performer. On the held-out test set it achieved a coefficient of determination, or R², of 0.7260, a root mean square error of 0.0734, and a mean absolute error of 0.0592, while also posting the highest cross-validation accuracy of the group with an R² of 0.802. For readers unfamiliar with these metrics, R² measures the fraction of variance in the outcomes that the model explains, with 1.0 being perfect, while the error metrics quantify the typical size of the model’s mistakes. A baseline R² above 0.72 on real-world driving data is respectable, but the authors were not satisfied.

The study’s most distinctive contribution lies in what came next: metaheuristic optimization. Metaheuristics are search algorithms inspired by natural processes, from the flocking of birds to the foraging of insects, that explore vast parameter spaces far more efficiently than brute-force grid searches. Chen and Zhou deployed three such algorithms to fine-tune the CatBoost model’s hyperparameters, the internal settings that govern how aggressively the model learns and how it balances bias against variance. The winner was the Horned Lizard Optimization algorithm, a relatively new nature-inspired method, which when coupled with CatBoost produced the study’s best-performing framework, dubbed HLO-CatBoost.

The improvements were substantial. The optimized HLO-CatBoost framework achieved a test R² of 0.8408, with the root mean square error falling to 0.0559 and the mean absolute error to 0.0434. In other words, hyperparameter tuning by the horned lizard algorithm lifted the explained variance by more than eleven percentage points while cutting the typical prediction error by roughly a quarter compared with the untuned baseline. That gap matters in practice: a battery management system that predicts remaining charge and range more accurately can make smarter decisions about when to heat the cabin, when to limit power, and how to keep cells inside their optimal temperature window, directly translating into extra kilometers per charge and slower long-term degradation.

Crucially, the authors did not stop at headline accuracy numbers. They subjected their framework to a battery of further analyses designed to probe its robustness and trustworthiness. Environmental scenario testing examined how the model behaves across different ambient conditions, a critical check for a system intended to operate from freezing winters to hot summers. Uncertainty analysis quantified the confidence one can place in individual predictions. Computational efficiency assessments confirmed that the framework is fast enough for practical deployment rather than being a laboratory curiosity. A Williams plot, a standard diagnostic from chemometric modeling, was used to identify outliers and define the model’s applicability domain, flagging any predictions that fall outside the territory the model learned from. Finally, feature importance analysis opened the black box, revealing which input variables most strongly drive the predictions of final state of charge and travel distance.

This emphasis on interpretability and validation distinguishes the work from much of the machine learning literature applied to batteries, where a single impressive accuracy figure is often the whole story. Battery thermal management systems, or BTMSs, are safety-critical components: they must prevent thermal runaway, a catastrophic failure mode in which a cell overheats and triggers a chain reaction, while simultaneously squeezing out every possible increment of efficiency. A model that cannot explain itself, or that fails unpredictably at the edges of its training distribution, is a liability in that context. By combining rigorous cross-validation, outlier diagnostics, scenario analysis, and feature attribution, Chen and Zhou have assembled the kind of evidence package that engineers and regulators need before trusting an intelligent management system with a live vehicle.

The broader implications reach across the electric vehicle industry. Range anxiety remains one of the most cited barriers to EV adoption, and a meaningful share of that anxiety stems from the unpredictability of range in extreme weather, when heating and cooling loads can slash advertised figures. Machine learning frameworks trained on genuine driving data, and sharpened by metaheuristic optimization, offer a path toward management systems that anticipate those losses and compensate for them, whether by pre-heating the battery while the car is still plugged in, by harvesting braking energy for the cabin, or by modulating thermal loads in real time. The study’s authors, who collected and analyzed the data and evaluated and edited the manuscript, report no competing interests, and their work was supported by research programs at Chongqing institutions along with technical support from the marketing department of Chongqing Changan New Energy Automobile. As electric fleets grow and software-defined vehicles become the norm, the marriage of real-world data, gradient boosting, and nature-inspired optimization may well become a standard ingredient in the batteries that move us, turning every braking event and every degree of ambient temperature into information that helps a car go farther on the same charge.

Subject of Research: Machine learning optimization of electric vehicle battery thermal management using real-world driving data

Article Title: Optimizing automotive battery thermal regulation strategies for enhanced efficiency in real driving cycles

Article References: Chen, B., & Zhou, C. (2026). Optimizing automotive battery thermal regulation strategies for enhanced efficiency in real driving cycles. Ionics. https://doi.org/10.1007/s11581-026-07474-3

Image Credits: AI Generated

DOI: 10.1007/s11581-026-07474-3

Keywords: electric vehicles, battery thermal management, real driving cycles, CatBoost, Horned Lizard Optimization, machine learning, state of charge, regenerative braking, BMW i3, hyperparameter tuning, range prediction, battery degradation

The study, published in Neural Computing and Applications, was conducted by R. J. Kuo and Yun-Tung Ko, who set out to tackle a problem that has long resisted conventional approaches. Classical statistical tools such as Auto-Regressive Moving Average models assume that the underlying data series behaves in a stable, consistent manner, an assumption that volatile wind speeds routinely violate. Deep learning models like Long Short-Term Memory networks, or LSTMs, have fared far better because their gating mechanisms allow information to persist across long stretches of time, but even these architectures have limits when faced with the non-stationary, noisy signals that characterize real wind farms.

The researchers’ solution layers three innovations on top of one another. First, they applied Variational Mode Decomposition, a signal-processing technique that breaks a turbulent wind power series into several narrow, smoother sub-components called intrinsic mode functions. Unlike earlier decomposition methods such as Empirical Mode Decomposition, VMD manages the bandwidth of each mode and incorporates Wiener filtering, which suppresses noise and avoids the spectral aliasing that can corrupt downstream predictions. Kuo and Ko did not simply accept default settings: they systematically selected the number of modes by minimizing an index of orthogonality, ensuring the decomposed components did not overlap, and chose the penalty factor by checking how faithfully the sum of the modes reconstructed the original signal.

Second, the team built a forecasting engine that fuses two complementary deep learning architectures. The first is a Temporal Convolutional Network, a convolutional design that uses dilated causal convolutions stacked in residual blocks. The dilation structure lets the network look far back into the historical record without piling on depth, and because convolutions can be computed in parallel, TCNs train faster and use less memory than recurrent alternatives. The second component is a Nested LSTM, a newer variant in which the memory cell of a standard LSTM is itself replaced by another full LSTM cell. This hierarchy allows the network to selectively access deeper memory, retaining and organizing information over extended periods, which is precisely the kind of capability needed to track slow-moving weather patterns superimposed on rapid fluctuations.

Rather than choosing between these two architectures, the researchers combined them. The TCN extracts local temporal features through its convolutional operations while the NLSTM captures long-range sequential dependencies. The output features from both networks are concatenated and passed through a single-layer artificial neural network that acts as a fusion layer, enabling nonlinear interaction between the two representations before the final forecast is issued. Earlier work had suggested that simply bolting a TCN onto an LSTM does not fully exploit what the hidden layers learn; the shared fusion layer in the new design was conceived specifically to address that shortcoming. The model also ingests weather variables alongside power data, with features such as wind speed, temperature, and solar radiation screened by mutual information scores that capture both linear and nonlinear relationships with the target series.

The third innovation concerns how the model is tuned. Deep architectures are notoriously sensitive to their hyperparameters, and the population-based optimizers commonly used to set them, including the Genetic Algorithm, can get trapped in local minima or converge prematurely. The authors turned to the Driving Training-Based Optimization algorithm, a recently proposed metaheuristic inspired by how student drivers learn: first from an instructor, then by imitating skilled drivers, and finally through personal practice. In each iteration, the algorithm evaluates candidate solutions, promotes the best performers to the role of trainers, and lets the rest of the population learn from them, with the pool of trainers gradually shrinking as the search progresses.

Kuo and Ko’s key twist was to inject mutation into this driving metaphor. Borrowing from evolutionary computation, they added a mixed mutation strategy in which Gaussian, Cauchy, power, or single-point mutations are applied to candidate solutions with equal probability. The purpose is to maintain diversity within the population and give the search a way to escape local optima, especially in later iterations when a standard optimizer would settle into a rut. The single-point mutation, in particular, alters only one component of a candidate’s position, echoing the idea that a learner might refine a single driving skill rather than overhauling everything at once. The objective function guiding the whole process was simple: minimize the Root Mean Square Error between forecast and actual wind power.

To test the framework, the team used five years of daily data from 2018 to 2022, covering 1,826 days from two European electricity bidding zones, DK2 and DE. Power generation figures came from the ENTSO-E Transparency Platform and weather data from Visual Crossing. After cleaning missing values, converting wind direction to radians to eliminate the discontinuity between 0 and 360 degrees, and normalizing all features to the zero-to-one range, the researchers split the data so that four years trained the models and the final year tested them. Every experiment was repeated thirty times to account for the natural variability of neural network training, and the hybrid TCN-NLSTM consistently produced lower average errors and smaller standard deviations than standalone TCN, NLSTM, or ARMA models.

The optimizer comparisons proved equally decisive. When the same hybrid architecture was paired with the Genetic Algorithm, the original DTBO, and the mutation-enhanced version, the improved DTBO delivered the lowest RMSE on both datasets with the greatest stability across repeated runs. Convergence curves showed DTBO-based methods improving rapidly at first and continuing to refine after ten iterations, with the enhanced variant exhibiting a characteristic zigzag pattern that signaled ongoing exploration rather than stagnation. Statistical testing backed up the visual evidence: a Kruskal-Wallis test confirmed significant differences among the compared models, and pairwise Mann-Whitney U-tests showed the hybrid architecture significantly outperforming ARMA, NLSTM, and TCN, with all p-values below 0.05. The gap between DTBO and its mutation-enhanced cousin, however, proved dependent on the dataset at hand.

The authors are candid about the framework’s limits. The current study forecasts only one day ahead with a fixed input window, and the fused architecture carries more computational weight than a single network, which could complicate deployment in resource-constrained, real-time settings. Some parameters were also fixed from prior literature or pilot experiments rather than optimized adaptively. Still, the consistent performance across two independent real-world datasets suggests the approach could find practical use in wind energy management and broader renewable integration efforts. As wind and solar are expected to account for roughly 95 percent of capacity expansion among clean energy sources according to the International Energy Agency, tools that shrink the fog of uncertainty around generation are likely to become as vital to the grid as the turbines themselves.

Subject of Research: Hybrid machine learning and optimization for short-term wind power forecasting

Article Title: Using hybrid machine learning model with an improved driving training-based optimization algorithm for wind power forecasting

Article References: Kuo, R. J., & Ko, Y.-T. (2026). Using hybrid machine learning model with an improved driving training-based optimization algorithm for wind power forecasting. Neural Computing and Applications, 38(17), Article 723. https://doi.org/10.1007/s00521-026-12401-8

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12401-8

Keywords: wind power forecasting, machine learning, Temporal Convolutional Network, Nested LSTM, Variational Mode Decomposition, Driving Training-Based Optimization, metaheuristics, mutation operator, renewable energy, hyperparameter tuning, deep learning, smart grids

NV4GP, short for Nested Validation for Genomic Prediction, was developed by Paolo Vitale of the International Maize and Wheat Improvement Centre, known as CIMMYT. The software is written entirely in Python and distributed under the permissive MIT licence through GitHub, with a permanent code archive on Zenodo. Its designers describe it as a user-friendly, efficient and reproducible framework for nested cross-validation-based model evaluation in genomic prediction, and its interface guides users through three sequential modules without requiring any command-line interaction or scripting. The motivation, the author explains, is that graduate students and early-career researchers entering plant breeding, biometrics or quantitative genetics often struggle to translate theoretical knowledge into practice because of limited coding experience, and accessible software can play a critical role in easing that transition.

The statistical heart of the software is its handling of hyperparameter tuning, a step that strongly influences model performance and generalization but is frequently mishandled in applied studies. Many prediction models, particularly non-parametric and artificial intelligence-based approaches, include one or more hyperparameters that are not estimated during model training. Nested cross-validation addresses this by defining an outer loop of conventional k-fold cross-validation to assess model performance, and then, within each outer-loop training set, running an inner-loop cross-validation that evaluates every combination of hyperparameters in a user-defined search grid. The combination with the lowest prediction error or the highest predictability is selected and used to fit the corresponding outer-loop model, ensuring unbiased performance estimates by preventing information leakage between model tuning and evaluation.

In an independent validation scenario, the software instead performs an internal k-fold cross-validation on the training population to select the best hyperparameters, then refits the model on the full training set and tests it on a genuinely separate dataset. Predictability is reported as Pearson’s correlation between predicted and observed values in the independent set. This strategy mirrors the real-world breeding context, in which breeders must use a training population to predict the performance of an unknown, genetically related breeding population. Independent validation has historically been under-applied in the genomic prediction literature, despite offering a fairer basis for model comparison than naive cross-validation schemes that inadvertently share information between training and testing data.

NV4GP implements fifteen prediction models spanning parametric, non-parametric and ensemble categories. The parametric set includes ridge regression, Bayesian ridge and LASSO, while the non-parametric range covers kernel ridge regression, support vector regression, elastic net, stochastic gradient descent, partial least squares, nearest neighbours, Gaussian process regression, decision trees, random forests, gradient boosting and multi-layer perceptrons, capped by a voting regressor ensemble. Each model carries its own hyperparameter grid; the multi-layer perceptron alone exposes twenty-one tunable settings, and gradient boosting and stochastic gradient descent each expose seventeen. All implementations draw on custom functions and the widely used Scikit-learn library, and a benchmark GBLUP model can be run in R through the BGLR package for comparison purposes.

Before any modelling begins, a marker filtering module lets users upload genotype data in numeric matrix or HapMap format, following IUPAC nucleotide nomenclature where applicable, and filter markers and genotypes by missing data, minor allele frequency and heterozygosity. Optional imputation by mean or major allele is available, and a detailed log reports exactly how many markers were removed by each filter and the per-marker quality statistics. In the validation module, users upload fully imputed marker matrices and phenotype files, which may be unbalanced because the software internally matches line identifiers and restricts analysis to common genotypes. Users specify the target trait, an optional logarithmic transformation, the hyperparameter grid, the number of cycles and folds for the outer and inner loops, a random seed for reproducibility, and the selection criterion: mean absolute error, mean squared error or predictability.

To demonstrate the software, the author ran five cycles of five-fold nested cross-validation, with four inner folds, on the wheat599 dataset of 599 individuals, 1,279 markers and four grain yield phenotypes recorded across environments. Predictability varied substantially across models: for the first target variable, support vector regression reached roughly 0.6 while decision trees scored 0.0; for another, the classical GBLUP benchmark led with values near 0.5, while Gaussian process regression lagged furthest behind; for a fourth, kernel ridge regression exceeded 0.5. Running times were equally revealing, ranging from about fifteen seconds for ridge regression to nearly twelve hours for the multi-layer perceptron on a conventional 64-bit laptop with 32 gigabytes of RAM. In an independent validation using published wheat data from two consecutive years and two simulated irrigation environments, predictability rankings shifted markedly between environments, ranging from slightly negative values for stochastic gradient descent to 0.32 for support vector regression in one environment and 0.16 for the multi-layer perceptron in the other.

The software was benchmarked against the GBLUP implementation in the BGLR R package using identical datasets and validation strategies, and produced comparable predictability estimates, supporting its correctness. Reproducibility is enforced through user-defined random seeds that fix the random state across all model components and fold partitioning, and the software was independently tested on Linux, Windows and MacOS, yielding identical results. Input validation checks flag common errors, such as misformatted marker files or incompatible cross-validation partitioning, before execution, returning informative error messages rather than silent failures. A result summary module processes output files and generates bar plots with standard error bars for test-set predictability, allowing direct visual comparison across models without additional data harmonization.

Compared with existing tools such as BGLR, rrBLUP, sommer, MegaLMM, ShinyGS, CHiDO and GS4PB, NV4GP is distinctive in combining a graphical interface with true nested cross-validation and independent validation, systematic grid-based hyperparameter tuning, and a broad roster of machine learning and deep learning models. The limitations are candidly acknowledged: exhaustive grid searches become computationally demanding as models, hyperparameter combinations or folds multiply, memory demands grow with large marker matrices, and the current version supports only a single trait or environment per run, deliberately excluding multi-environment and multi-trait modelling to keep the software usable on standard laptop hardware. The recommended ceiling for a conventional laptop is a search grid below fifty hyperparameter combinations on datasets of hundreds of lines with a few thousand markers.

The implications extend beyond convenience. By collapsing what previously required substantial bespoke programming into a uniform graphical workflow, and by standardizing reported metrics including training and test predictability, mean squared and absolute errors, percentage errors and the gap between training and test performance, NV4GP makes cross-study and cross-model comparisons of machine learning methods in genomic prediction more reproducible and more accessible. For breeders, the practical consequence is a shift in how model-selection decisions are made: rather than defaulting to a single familiar model because alternatives are impractical to implement, users can rapidly screen multiple model families and choose the empirically best performer for their specific trait and population, embedding statistically rigorous, leakage-free validation into routine genomic selection pipelines.

Subject of Research: A Python software framework for nested cross-validation and independent validation in genomic prediction for breeding

Article Title: NV4GP: Nested validation for genomic prediction

Article References: Vitale, P. (2026). NV4GP: Nested validation for genomic prediction. SoftwareX, 36, Article 103021. https://doi.org/10.1016/j.softx.2026.103021

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103021

Keywords: genomic prediction, genomic selection, nested cross-validation, hyperparameter tuning, machine learning, plant breeding, quantitative genetics, NV4GP, CIMMYT, wheat, Python software, model benchmarking

The core of the proposal is deceptively simple in structure. Experts first specify a chain of progressively more demanding goals, ordered in a Guttman-like sequence from a trivial baseline objective to a final, fully realized goal. Each goal on the chain is paired with a detector—a logical clause or standardized test—that can be checked for every individual service request arriving at a social agent, whether that agent is a healthcare provider, an education service, or any other organization responding to citizens’ needs. Because the goals are strictly ordered, any instance can be tested forward along the chain until it fails, yielding an ordinal level score between zero and the maximum L.

What transforms this ordinal skeleton into a genuine measuring instrument is the assignment of gain scores—non-negative value increments attached to each transition along the goal chain, possibly conditional on the state of the service requester. In a care-assistance example developed in the paper, the goals range from the ability to stand up and move inside a room, through personal hygiene, toileting, and eating, up to cooking, obtaining food from outside, and communicating with the wider world. A client who advances further along this chain accumulates the corresponding sum of gain scores, so the raw levels are re-quantified into a meaningful value scale. An evaluation functional then aggregates these quantified levels across the entire mass of service instances handled by an agent, weighted by the share of clients who clear each rung, producing a single social-value index. Notably, the author observes that this functional has the mathematical form of a Yaari-Quiggin value functional from rank-dependent expected utility theory, although the interpretation here departs from standard utility-based readings.

The methodological heart of the study lies in answering where those gain scores should come from. Rather than leaving them to ad hoc expert judgment, D’Epifanio proposes extracting them automatically from a reference training-data set associated with a chosen benchmark agent—a real organization or, conceptually, a virtual ‘good practice’ whose behavior defines the standard against which all others are calibrated. The criterion governing this extraction is an intuitive principle the author calls intrinsic worthiness: the greater the resistance—or risk of failure—encountered in reaching the next goal, given that the previous one has been achieved, the greater the value increment that should be credited to an instance that does succeed. In probabilistic terms, the gain attached to each step equals a function of the conditional probability of failing that goal, estimated from the reference data.

A worked example illustrates the mechanics concretely. Focusing on elderly male clients aged 75 to 85 in a training data table, the empirical failure probabilities rise steeply along the chain: roughly 0.069 for the first transition, about 0.20 for the second, then 0.83 and 0.90 for the hardest final steps. These numbers directly quantify how much harder each successive achievement is for that population, and therefore how much more value should be assigned to clients who manage it. A family of logistic models can further link the failure probabilities to client characteristics such as sex and age class, allowing the quantified scale to shift appropriately across different conditions and strata of the served population.

The most technically ambitious portion of the paper develops an advanced, model-based pseudo-Bayesian approach within a Dirichlet-multinomial framework. The training data, organized as a contingency table crossing client conditions with performance levels, are modeled by conditionally independent multinomial distributions governed by latent probability profiles, which are themselves given Dirichlet priors parameterized by expectation profiles and vagueness descriptors. Hidden Markov-type transition probabilities, representing the chance of failing each goal, are embedded in these expectation parameters and driven by a logistic regression engine—essentially a hidden layer connecting client attributes to underlying performance dynamics—regulated by an extended set of hyperparameters that the author terms hyper-hyperparameters.

Tuning those hyper-hyperparameters is where differential geometry enters the picture. The full space of Dirichlet distributions is viewed as a container manifold, while the distributions constrained by the hidden regression engine form a nested submanifold. A pseudo-Bayesian operator of updating, computable in closed form, acts as a vector field on this manifold, expressing how the training data would revise any given distribution. The tuning criterion is an orthogonality principle: the desired configuration of hyperparameters is the one at which the updating gradient, evaluated at the constrained distribution, lies entirely orthogonal to the tangent space of the submanifold. Intuitively, this is the point where the training data add no fresh information beyond what is already structurally encoded in the modeling constraints themselves.

The author connects this geometric condition to a minimum information principle, echoing ideas from Bayesian conditionalization theory: the less a pre-formatted knowledge representation is updated by current data, the more the information carried by those data is already intrinsically contained within it. Once the optimal hyperparameter profile is found, the hidden failure probabilities can be recovered backward, the standardized gain scores extracted, and conditional quantified scales constructed for every client state—implemented computationally with an algorithm written in the R language. The resulting indexes can then be normalized to the interval from the ideally worst to the ideally best virtual agent, aggregated across multiple social strata with designer-chosen weights, and applied to rank any number of competing providers conditionally on circumstance.

The implications extend well beyond one methodological exercise. As automated and artificially intelligent agents increasingly participate in the delivery of public services, transparent and replicable evaluation standards become urgent. The framework presented here offers a structured, potentially fully automatable pipeline: specify ordered goals, gather reference behavior data, invoke the intrinsic-worthiness criterion, tune the model geometrically, and emit standardized, condition-aware performance scores for any agent in the class. Because the entire calibration is anchored to explicit design specifications and an auditable data set, the resulting indexes carry an interpretable probabilistic meaning that conventional composite indicators often lack. The study thus contributes a principled bridge between the statistical machinery of modern machine learning and the enduring civic need to measure, compare, and improve the social value of public services.

Subject of Research: A probabilistic and geometric method for calibrating social performance indexes from training data

Article Title: Tuning Social Indexes on Training-Data

Article References: D’Epifanio, G. (2026). Tuning Social Indexes on Training-Data. Social Indicators Research, 184(2), Article 37. https://doi.org/10.1007/s11205-026-03923-8

Image Credits: AI Generated

DOI: 10.1007/s11205-026-03923-8

Keywords: social indicators, performance indexes, training data, Dirichlet-multinomial model, pseudo-Bayesian methods, hyperparameter tuning, information geometry, Guttman scale, intrinsic worthiness, social service evaluation, machine learning, statistical modeling