Papers

This study examines emission interdependencies within the waste sector and between waste and other economic domains, energy, transport, water, and communications, using Life Cycle Assessment (LCA) methodologies with emphasis on process-based and hybrid approaches. The analysis employs the EXIOBASE multi-regional input-output database. Statistical correlation is applied to identify commonalities and divergences in emission drivers. Results indicate strong correlations among certain landfill waste streams, particularly organic-based materials, as well as links between plastics and wood, suggesting shared decomposition or management pathways. Moderate correlations appear in cross-sectoral interactions, especially between waste, energy, and transport, while weak correlations are observed in processes with distinct treatment mechanisms, such as landfill versus wastewater. Methodologically, the study highlights challenges posed by high-dimensional datasets, including multicollinearity, which limit the effectiveness of linear models. These findings underscore the complexity of sectoral interdependencies and the need for both integrated and material-specific strategies. By demonstrating how systemic tools like EXIOBASE can capture these linkages, the research provides a foundation for constructing more robust emission models and informing targeted mitigation policies. The results contribute to advancing environmental accounting frameworks while offering practical insights for emission reduction planning across interconnected sectors.

We study whether causal risk mandates constructed from overlapping information blocks can be implemented by one self-financing portfolio that is optimal under pooled information. In an incomplete continuous Brownian market, signal-responsive exposures are projected onto traded Brownian directions and embedded as closed subspaces of a predictable Hilbert space. Each causal mandate is the attainable projection of an identified, baseline-centered intervention surface and is therefore fixed upstream of the allocator. A first exact decomposition separates non-traded structural response, incompatibility across local mandates, and distortion of the common traded book. A fusion operator corrects duplication of common risk directions. The main theorem shows that exact causal decentralization holds if and only if the pooled enlargement preserves reference martingales, the pooled optimum has no component outside the additive desk span, the local causal shadows are compatible with one common book, and that book equals the additive pooled optimum. Under immersion, log-growth regret separates into pooled-interaction loss and common-book distortion, while raw causal implementation error also contains non-attainability and incompatibility. No common benchmark can reduce incompatibility. We derive the unique allocation under a quadratic causal-mandate penalty and exact comparative statics in a two-desk model with shared, desk-specific, and pooled-interaction risk. Additional results cover joint interventions, representation stability, admissible coalitions, and a covariance-weighted continuous-semimartingale formulation. The conclusions are partial-equilibrium statements for continuous markets.

Price-level limit order book (L2) data reveal aggregate liquidity but not the ordered queue required by price–time priority. Passive-execution backtests can therefore depend on an unobserved cancellation-allocation rule even when observed prices, quantities, and trades are held fixed. We frame recovery of market-by-order histories from aggregate snapshots as a conditional partial identification problem: multiple histories can reproduce the same aggregate path. Holding that path, reconciled market removals, latent order partitions, and additions fixed, our path-preserving compiler varies only cancellation allocation among front, quantity-weighted-random, and back rules. Within this compiler class, we establish front–back fill ordering for a virtual tagged order during one touch-price spell. We study seven months of synchronized 2025 Tokyo Stock Exchange data for two instruments with different trading activity: RIC 1301.T and RIC 7911.T. Ten-level L2 snapshots provide book states, while L1 trades permit inference of market removals and resting side through reconciliation. Each instrument contributes 1,080 matched five-minute episodes over the same 18 held-out trading days. The aggressive benchmark is invariant across FIFO realizations, but passive execution is sensitive to the cancellation rule. For 1301.T, front rather than back cancellation raises preterminal completion by 8.01 percentage points and reduces implementation shortfall by 1.010 bps. For 7911.T, the corresponding differences are 7.39 percentage points and 0.384 bps. Thus, observationally equivalent aggregate-book paths can imply economically different passive-execution outcomes. Execution policies evaluated from aggregate data should be accompanied by FIFO sensitivity analysis rather than reported as single-point estimates based on an unobservable queue assumption.

We develop a diffusion-model framework for dynamic implied-volatility surface generation and evaluate its economic usefulness through data-driven hedging. The framework consists of two models. AD-Seq-Vol jointly learns the conditional evolution of the underlying asset return and the high-dimensional implied-volatility surface, generating adapted multi-period scenarios by sequentially updating the realized market history. AD-Seq-Vol-FT further incorporates option-market structure through post-training penalties for violations of static no-arbitrage conditions. Using daily SPX option data from 2000 to 2023, we show that the proposed models generate coherent surface trajectories while capturing both cross-sectional and temporal dependence. AD-Seq-Vol produces fewer and less severe static-arbitrage violations than the training data and the GAN-based benchmark, while AD-Seq-Vol-FT reduces these violations to nearly zero. We then integrate the generated conditional scenarios into an optimization-based hedging framework. Compared to a range of classical and data-driven benchmarks, the diffusion-based hedges maintain tracking errors near zero, substantially reduce tail risk, and exhibit particularly stable performance during the COVID-19 market disruption. These results establish Adaptive Sequential Diffusion Models as a promising class of market-consistent and economically useful financial scenario generators. Our code is available at: https://github.com/yinbinhan/volatility-surface-simulation.

Empirical researchers often use binary classification trees to identify subgroups at risk of adverse outcomes. We compare two classification tree algorithms in this policy-targeting context: classical classification and regression trees (CART) and the maximizing-distance final-split approach (MDFS). We establish a theoretical setting in which MDFS identifies a high-risk subgroup while CART identifies none. Applied to rural hospital closure data, MDFS targets a high-closure-probability subgroup: for-profit hospitals with more than 63 days in accounts receivable. CART misses this subgroup. To facilitate further application of these classification tree algorithms, we provide targetree package in Python, R, and Stata.

Whether medical technology reduces or increases demand for professional care is central to understanding its effects on healthcare utilization and costs. We study automated insulin delivery (AID) adoption among 1,608 adults with type 1 diabetes treated in four specialist clinics of the Italian National Health Service, 283 of whom adopt AID. Because adoption is clinically targeted and staggered, we combine risk-set coarsened exact matching with a staggered event-study design. Matching retains 181 adopters with comparable untreated controls. We further adjust for differential pre-adoption trends by extrapolating the untreated trajectory in event time. After adjustment, AID adoption is associated with greater routine outpatient engagement. The estimated effect on the probability of a diabetologist visit is 16.3 percentage points one semester after adoption and 39.3 percentage points four semesters after adoption, relative to a 55.2% visit probability in the semester before adoption. HbA1c testing and a broader process-of-care index show similar positive patterns, although the evidence is less precise. The visit estimates remain sensitive to the trend-extrapolation assumption and should therefore be interpreted conditionally on that restriction. The findings are consistent with a task-based view in which automation substitutes for routine dosing decisions while complementing clinical labor through interpretive, adjustment, and supervisory tasks. Patient-facing automation may therefore reorganize rather than simply reduce healthcare use.

The favorite–longshot bias (FLB) is one of the most persistent return patterns in betting markets, but its prevalence in modern prediction markets and its relation to recurrent longshot demand remain unclear. We study these questions in Polymarket, a large and increasingly important prediction market, using 588 million trades by 2.48 million accounts. In observed transaction flows, purchases below 10 cents lose 19.3 cents per dollar, while purchases at or above 90 cents earn 0.83 cents. How contracts are grouped changes the result: longshots lose 6.3 cents per dollar when each contract is weighted equally but gain 4.1 cents when related contracts are first grouped by parent event; the two-sided pattern is robust in Crypto and Politics but surprisingly absent in Sports. The top decile of accounts with the strongest past tendency to buy longshots accounts for 26.6 percent of next-month low-price purchases, but earns similar returns to others. The top decile of accounts with the strongest past tendency to buy favorites accounts for 15.1 percent of next-month high-price purchases but earns less than others. These findings provide evidence of an aggregate FLB on Polymarket, but its magnitude depends on how contracts are grouped.

Stock price prediction remains challenging due to the non-stationary and noisy nature of financial time series. Existing deep learning models often rely on rigid decision-level fusion, ad hoc hyperparameter tuning, and compressed final-layer outputs, causing information loss, overfitting, and limited cross-market generalization. We propose VertiFuseX, a hybrid LSTM architecture using penultimate-layer vertical fusion of multi-scale temporal representations. VertiFuseX stacks and reweights penultimate features from LSTM, Bi-LSTM, and St-LSTM branches, integrates a parallel DNN stream, and jointly optimizes all components via backpropagation under a fixed hyperparameter configuration. This preserves richer intermediate temporal information across scales. Evaluated on 15 years (2010-2024) of closing prices from 10 global equity indices using strict chronological out-of-sample testing with the final 365 trading days held out, VertiFuseX achieves 30-54% MAPE reductions and over 40% improvements in MAE and RMSE versus LSTM-based baselines, and outperforms seven state-of-the-art models across 33 metric-dataset comparisons. Ablation studies confirm penultimate-layer fusion drives these gains over final-layer fusion and decision-level ensembling. Gradient-based saliency analysis shows consistent emphasis on mid-range dependencies at lags 9-15 days. Economic validation via algorithmic trading simulation under extreme market regimes shows reduced maximum drawdowns and superior risk-adjusted returns. With 675k parameters, a 2.6 MB memory footprint, and 1.5 ms/sample inference latency, VertiFuseX offers a lightweight, interpretable, deployment-ready framework for robust financial forecasting.

In this note, we propose a general structural approach to model a default time $\tau$ as the first-passage time (FPT) of a (“firm-value”) process $S$ below a (“debt”) barrier $K$ that comply with a pre-specified survival probability curve $G(t)=\Pr(\tau>t)$. Following an idea of Mbaye and Vrins (Mathematical Finance, 2022) applied to reduced-form models, our approach consists in two steps: choose a latent FPT model driven by a barrier $\tilde{K}$ and process $\tilde{S}$, and time-change those using a deterministic clock $\Theta$ to get $K_t=\tilde{K}_{\Theta(t)}$ and $S_t:=\tilde{S}_{\Theta(t)}$, leading to the final FTP model $(K,S,\Theta)$. As the market curve $G$ and the latent model $(\tilde{K},\tilde{S})$ are assumed to be given, the calibration step simply consists in finding the clock $\Theta$ such that the distribution of the FPT of $S$ below $K$ coincides with the survival curve $G$. We show that this is achievable for a broad class of specified curves $G$ and latent FTP models. The calibration amounts to a simple inversion of a function, which is almost immediate provided that the latent model is tractable enough. In particular, we show that the AT1P model of Brigo, Morini and Tarenghi - which is able to reproduce a broad range of CDS term-structures - can be regarded as the FPT of a time-changed drifted Brownian motion to a constant barrier: $\tilde{V}_t=\mu t+W_t$ and $\tilde{K}_t=k<0$. This connection offers an elegant interpretation for the instantaneous volatility function featured in AT1P and yields an immediate calibration of the latter to perfectly match a target survival curve.

We give a two-period counterexample to the absence of arbitrage for the posterior-weighted pricing rule in Estimate nothing by Duembgen and Rogers. Both physical models have strictly positive transition densities, and each model is equipped with an equivalent martingale measure. Nevertheless, the mixed price of a single derivative falls deterministically from $5/2$ to $2$ between two trading dates. If these prices are tradable, shorting the derivative and closing the position one period later yields a certain profit. A finite-state appendix also illustrates the failure of recursive consistency.

An ensemble of roughly 3,000 signals over 20 assets was reported to have approximately 90% correlation with the leading component of the asset-space return structure. Does having about 158 signals per available linear dimension explain that alignment? The population answer depends on the research process's design distribution and its relation to returns: crowding alone imposes neither a nonzero mean nor agreement with a principal component. The population theory developed here distinguishes four objects: the equal-weight signal, signal principal components, equal-weight profit and loss (PnL), and PnL principal components. The return operator in the motivating observation is a separate object. Independent libraries converge to their design mean; exchangeable libraries can retain a random conditional mean. Cross-sectional signals over $d$ assets, once demeaned and normalized, lie on the unit sphere $S^{q-1}$ of a $q$-dimensional space, $q=d-1$. Under axial symmetry, their nonzero mean is signal-cloud PC1 exactly when its longitudinal second moment exceeds $1/q$, the isotropic energy share. Residual-and-gap bounds quantify approximate alignment. Combining design weights into signals contracts the tangent of their angle to PC1 to at most $\sqrt{\lambda_2/\lambda_1}$ times its value, where $\lambda_1>\lambda_2$ are the leading signal-kernel eigenvalues; the factor is sharp. A target-aligned frame separates transverse signal geometry, which PnL discards, from dispersion weighting and temporal centering, which also change the spectrum. A reproducible synthetic example illustrates the geometric threshold, and a proposed empirical program separates library growth from limited-history estimation. No market-data empirical results are presented.

Using a balanced panel of 129 countries from 2000 to 2024, we estimate the macroeconomic costs of sanctions on Russia. To that end, we employ a novel two-wave common-correlated-effects imputation design that separates the 2014 sanctions from the 2022 escalation and accounts for heterogeneous exposure to global shocks. We find the main boomerang effect of these sanctions operates through prices rather than output. Specifically, relative to 2019-2021, we find a significant post-2022 increase in prices among sanctioning countries that remains robust under alternative counterfactuals, reference windows, and inference procedures. Conversely, we find no robust evidence of aggregate GDP-per-capita losses.

Mean–variance portfolio choice takes the conditioning information as given and optimizes over weights, so two errors about that information pass into the portfolio unseen: using variables unavailable at the decision time and treating common variation as idiosyncratic. We make the conditioning information a decision variable subject to hard admissibility constraints declared in advance: availability, a no-arbitrage-preserving enlargement of the decision filtration, statistical separation and, for interventional claims, invariance across declared regimes. A lexicographic admissibility order in which no risk–return quantity enters selects an optimal information class, and the classical problem is solved inside it. We prove existence, invariance under recodings, and a value-of-admissible-information theorem whose failure conditions show that selection by decision loss admits look-ahead information whenever present; the two-stage solution and the joint envelope are the minimal elements of two orders on one set. Under exact separation the diversifiable part of risk is a property of the admissible class and is interventionally stable only under interventional admissibility; requiring causal identification carries an explicit oracle price traded against search complexity. The estimator is consistent with second-order regret. On market data with 127 candidate drivers the condition is attainable on individual equities, where enlarging the search reduces the defect at a measurable rate, and not on pre-diversified portfolios, where it is nearly invariant to the search because the residual dependence is the common factor itself. Where it fails, the covariance still yields minimum-variance portfolios that match shrinkage at markedly lower turnover while discarding a measurable part of their risk, which an exact decomposition attributes to the residual share, breadth and average residual correlation.

Commodity futures are shaped by harvest cycles, weather shocks, storage conditions, and seasonal demand, but it remains unclear whether recurring patterns yield robust out-of-sample trading profits. Existing research documents return seasonality in commodity futures as well as more complex seasonal structure, while leaving less evidence on how alternative seasonal models compare under common implementation constraints. This article compares dummy-variable regression (DVR), Singular Spectrum Analysis (SSA), and robust low-rank SSA (RLSSA) within a unified trading framework, including a volatility-normalised specification. Using monthly delivery-avoidance returns for 15 liquid commodity futures, the models are estimated on rolling ten-year windows and evaluated from 2016 to 2024 with transaction costs, an equal-weight long benchmark, and Maximum Entropy Bootstrap (MEB) assessment. Across 500 MEB paths, the benchmark has the strongest average full-period profile, with a cumulative return of 16.81%, a Sharpe ratio of 0.191, and a maximum drawdown of -0.414. Classical SSA has the strongest average model outcomes, but negative median cumulative returns and deep drawdowns indicate substantial path sensitivity. Volatility normalisation reduces the average DVR short loss but generally weakens SSA-based portfolios. None of the 18 approximate paired MEB Sharpe tests rejects after within-family Holm adjustment. Because MEB preserves each contract's temporal rank ordering, the assessment is conditional on observed timing rather than a timing-randomised seasonal null. The evidence does not establish robust benchmark outperformance and shows that model performance varies materially across market subperiods.

We develop Entropic Value-at-Risk (EVaR) parity for tempered stable returns. EVaR-based inverse risk parity (IRP) and equal risk contribution (ERC) portfolios are constructed using multivariate normal tempered stable models and independent component analysis with tempered stable components. We derive the corresponding asset-level EVaR and EVaR-deviation contributions and use the latter to separate the fitted location term from EVaR risk contributions. Under Gaussian returns, EVaR-deviation IRP and ERC recover conventional volatility IRP and ERC weights. We evaluate the resulting portfolios in three investment universes. Empirically, EVaR-based ERC portfolios achieve positive Sharpe differences relative to equal weight across the universes.

Classical market-making strategies based on stochastic control, such as the Avellaneda-Stoikov and the Guéant-Lehalle-Fernandez-Tapia (GLFT) extension, provide closed-form quoting rules, but rest on assumptions that break down at realistic microstructure timescales. One of them is that order flow is stationary, while empirical evidence points to the existence of regimes, possibly associated with algorithmic execution of metaorders. In this case, existing methods provide negative PnL. In this paper, we develop a deep reinforcement-learning market maker (RLMM) - a Rainbow-style distributional DQN (C51) which is calibrated and tested in a zero-intelligence limit order book. We find that, in the stationary setting, RLMM outperforms GLFT across the entire observed risk-return frontier. The RLMM is more robust to flow asymmetry than GLFT, but, like any stationarily trained strategy, it still suffers large drawdowns from inventory saturation under persistent directional imbalance. Augmenting the state of RLMM with two auxiliary signals - a Bayesian online change-point filter over the directional flow bias and a queue-adjusted quote-exposure imbalance -restores profitability. A final scenario-bandit step that reweights low-return regime scenarios further improves performance under random-persistence and correlated-direction stress.

We prove that the skew stickiness ratio converges to two at short maturity under local volatility models. This appears to be the first rigorous proof of this limit for a general time-dependent local volatility function. As a by-product, we strengthen the one-half rule of the implied volatility skew by removing uniform ellipticity and global bounds on spatial derivatives of order at least two. The proof uses a first-order Watanabe expansion.

Modeling the joint distribution of extreme values in high-dimensional financial time series is challenging because extremes are sparse and locally extreme observations are not necessarily extreme relative to their full marginal distribution. To address this, we introduce a time-dependent network Hüsler-Reiss model in which market-informed adjacency matrices determine how strongly observations contribute to the estimation. We propose binary and weighted specifications, including the Joint Extremes Adjacency Matrix (JEAM) which combines information about individual extremeness with historical patterns of joint extreme movements. In the forecasting evaluation part, covering one-minute stock returns from three sectors of the S&P 100, JEAM achieves the best out-of-sample log scores for both tail directions; improving scores by 12.5-13.6% in the lower tail and 11.4-14.9% in the upper tail. The results show that incorporating market-informed network structures in the estimation, improves forecast evaluation of extremes across time series.

Constructing a confidence interval for the difference between two independent binomial proportions involves a nuisance direction that is not identified by the estimand. The one-sample Wilson score interval inverts a scalar score test, but has no direct bivariate analogue isolating the difference: inverting the joint normal approximation yields an elliptical region in the unit square, whereas the estimand \(p_1-p_2\) is one-dimensional. We define the Elliptically Optimal (EO) confidence interval as the range of \(p_1-p_2\) over this region and solve the resulting optimization problem in closed form, obtaining explicit bounds in six mutually exclusive and exhaustive cases. The solution admits a compact characterization: the EO interval is the score interval obtained by maximizing over the nuisance variance rather than estimating it. It is therefore the shortest interval obtained by projecting the elliptical region, and inherits its coverage guarantee. We derive the exact coverage excess, \(2[\Phi(z\mathcal{R})-\Phi(z)]\), where \(\mathcal{R}\) is the ratio of the least-favourable to the true standard deviation. The excess vanishes on an explicit line through the parameter space, is bounded by \(\alpha\), and is invariant under proportional scaling of the sample sizes. Exact enumeration of the binomial coverage shows that the Wald interval, whose variance estimator is downward biased by a factor \(1-1/n\) under balanced allocation, falls below nominal coverage almost everywhere. The EO interval never under-covers under the normal approximation and always yields admissible, non-degenerate bounds. Its price is over-coverage when both proportions are extreme, which we quantify exactly.

Decentralized exchanges record trading and liquidity provision on public blockchains, but empirical analysis requires interpreting these records and linking them to execution metadata. dexamine is a Python package that parses Uniswap v2 and v3 events on Ethereum. It converts transaction receipt logs into observations of trades and liquidity changes, with token quantities, pool state, transaction order, and gas information. The package separates data retrieval, contract metadata, protocol interpretation, and output construction. The repository provides recorded Ethereum responses and an offline reproducible example, and version 1 has been used to construct data for an empirical study of price discovery in decentralized markets.

Whether organisational separation of elective and emergency care improves physician productivity remains an open question. Most existing evidence relies on cross-sectional comparisons or volume-based outcomes that cannot isolate efficiency gains from input expansion or provider selection. In contrast, this paper exploits the staggered rollout of NHS England's surgical hub programme, a set of dedicated ring-fenced elective facilities introduced as part of its Elective Recovery Plan, to provide causal estimates of the effect of elective-emergency separation on physician productivity. Using this rollout to implement a heterogeneity-robust difference-in-differences design, we estimate the impact of hub adoption on physician productivity measured as cost-weighted elective output per unit of salary-weighted physician input. We find that hub adoption increases physician productivity by 14.5% relative to the no-hub counterfactual, accompanied by a 10-day (7.6%) reduction in average patient waiting times. The size of these gains depends on how completely elective care is insulated from emergency pathways. Standalone hubs situated on dedicated elective-only sites achieve relatively larger productivity gains than integrated hubs located within main acute hospital sites. Moreover, operating multiple hubs yields gains more than double the overall average, whereas a single hub shows no statistically significant effect. For health systems seeking to raise physician productivity and address elective backlogs, these findings suggest that how separation is implemented, not simply whether it is adopted, shapes the productivity gains it delivers.

We present a comprehensive description of the network of Italian Business-to-Business commercial relationships, based on the universe of electronic invoices collected by the Italian Tax Office. The firm-to-firm detail of the data is exploited to describe the distribution of the numbers of buyers (customers) and sellers (suppliers) per firm, the centrality of each firm in the production network, and the average distance between firms. We characterize for the first time the firm-to-firm network describing the Italian production system, and present its geographic and sectoral breakdown. The analysis reveals heavy tailed distributions for the numbers of buyers and sellers per firm, implying a scale-free structure of the network. The distributions of centrality values display heavy tails as well, indicating a strong concentration of importance in a relatively small number of firms. We estimate the tail exponents for all these distributions, finding in all cases lower exponents for downstream (buyer-side) distributions than for the upstream (seller-side) ones.

We investigate the strategic communication of narratives under model uncertainty. The sender has private information about the true data-generating process of publicly observable data. The receiver is uncertain about how to interpret the data, but aware of the sender's incentives to strategically provide interpretations (“narratives”). We theoretically show that the size of the conflict of interest between the sender and the receiver is a crucial determinant of equilibrium communication. In particular, the stronger the sender's bias, (i) the more senders exaggerate their information and (ii) the more receivers correct the sender's action recommendations. In a laboratory experiment, we find evidence in line with both predictions, suggesting that people in complex and uncertain environments take a narrator's strategic incentives into account. Additional analyses reveal that narrative likelihood does not drive receiver behavior and that narratives are, on average, slightly persuasive only when bias is low.

In many power systems, wind and solar generation increasingly often exceeds electricity demand. Curtailing renewable generation in those hours matters both for prices and for the physical stability of the grid. Turning off wind turbines and solar panels is technically easier than ramping down a large power station, yet support schemes often give renewables an economic incentive to keep producing at negative prices. This paper studies wind and solar energy in Germany. For each cohort of generators it estimates, hour by hour, the incentive implied by two decades of support policy. It then sets those incentives against observed behavior, using a new estimate of market-based curtailment built from reanalysis weather data. I find that in 2025, at prices below -50 EUR/MWh, almost all wind generators had an incentive to stop producing, but only half of them did. Solar is the opposite case: nearly two thirds of the potential had no incentive to curtail at all, mostly because it receives a feed-in tariff that shields it from wholesale prices. Of the exposed remainder, just over a fifth cut production. Low exposure and response rates inflate subsidy payments and make the power system harder to operate safely. I conclude that a further expansion of wind and solar requires them to respond to price signals.

Machine learning-based credit scoring is increasingly central to Peer-to-Peer (P2P) lending, yet its resilience to adversarial manipulation, where applicants strategically alter self-reported inputs to secure favourable decisions, remains poorly understood. Most adversarial-robustness evidence comes from image and text domains and evaluates a single attack against a matching defence, offering little guidance on how defences generalise across attack types in tabular credit data. We address this with a systematic train-test robustness benchmark on a large Lending Club subset, spanning three model families (logistic regression, a feed-forward neural network, and a transformer for tabular data) and four attacks confined to applicant-mutable features: Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Salt-and-Pepper (S&P) noise, and DeepFool, plus a mixed-attack regime. Across a full grid evaluated with stratified cross-validation, adversarial training sharply improves robustness against the attack it is trained on and transfers well within the gradient-based family, but transfers weakly to non-gradient corruption, so single-attack defences overstate real-world resilience. Mixed training delivers the most balanced robustness across heterogeneous attacks while preserving clean-test performance, supporting multi-attack stress testing in credit-model governance.

When a country's population declines, the aggregate economy appears to contract on the intensive margin: industrial diversity intact, every industry a little smaller. At the regional level, contraction is uneven and takes the extensive form: entire industries disappear, one after another. The relevant unit is the city: industries are nested by size - the hierarchy property of industrial location - each viable only above a minimum population. Necessity industries' thresholds bunch at the low end, so a city's industry count - and its jobs - is sharply concave in size (concavity on jobs). A modest loss pushes a small city below many thresholds at once; a large core sheds a few specialized industries, one at a time. Lost industries consolidate upward to the next city large enough to host them; for the worker it means a step down to a lower-paid local job. To recover that income, workers move up to the apex - the only city hosting the full industry range. Studying Japan - two decades ahead of the OECD, Tokyo at its apex - with worker-level panel data on the young workers who carry the migration, a wage regression in real, housing-inclusive wages identifies a Tokyo-bound migration incentive that varies by origin, following concavity on jobs.

Drug-drug interaction (DDI) errors pose serious risks to patient safety. Existing decision-support systems often require physicians to respond to alerts, disrupting workflows and contributing to high override rates. We examine whether a non-mandatory information intervention can reduce DDI errors and foster learning. Using a randomized field experiment with India's largest electronic medical record platform, we analyze 2.81 million prescriptions from 1,700 physicians using a difference-in-differences design. Treatment physicians received real-time information highlighting DDI errors without being required to respond, while control physicians received no such information. The intervention reduced DDI errors by 8.6%, corresponding to an estimated US$4.8 million in annual hospitalization cost savings and approximately 134 lives potentially saved. We identify two mechanisms: reactive correction, whereby physicians remove errors after they are flagged, and proactive learning, whereby they avoid errors before alerts occur. While early reductions are driven primarily by correction, physicians increasingly avoid errors over time. They also become less likely to repeat previously flagged errors and reduce new errors, suggesting that learning generalizes beyond specific drug pairs. The effects are consistent across physician types and do not compromise productivity or care quality. Our findings show that non-mandatory information interventions can improve patient safety through both immediate error correction and persistent, generalizable learning.

Signal correlation and PnL correlation are correlations over different index sets - across assets at each date versus across dates for scalar payoffs - and practitioners often treat the first as a proxy for the second. We give an exact decomposition that shows what that proxy sees and what it discards. We recall that at each date a normalized signal's projection onto the realized demeaned return direction is its realized cross-sectional Pearson IC, so that fixed-signal PnL is return dispersion times IC (Qian and Hua, 2004), and we identify the normalized similarity of two signals in the orthogonal complement as their partial correlation controlling for realized returns; the residual rotational freedom is an orthogonal gauge whose invariants are the transverse Gram matrix. Signal correlation therefore equals an uncentered IC cross-moment plus transverse similarity, while Pearson PnL correlation centers and dispersion-weights the IC series alone. Our main result is a non-identifiability theorem: absent constraints on transverse geometry, neither correlation bounds or orders the other, which sharpens the simulation finding of Sorensen, Qian, Schoen, and Hua (2004) into an exact statement. For normalized ensembles the transverse Gram matrix re-enters through the normalization denominator; the ex post optimal combination under Euclidean and general covariance-risk metrics is the classical multiple-correlation bound, and we show that the two rules agree for every longitudinal exposure if and only if the risk metric is isotropic on the signal span. The classical exact null law of partial correlation under isotropy, a synthetic mechanism illustration at weak-IC scales, and inference guidance under temporal dependence complete the treatment.

This paper develops a public-data framework for evaluating incremental bank lending when plausible climate scenarios imply different sector credit outcomes but defensible scenario probabilities are unavailable. The calibration combines 2025 Shared National Credit industry commitments, January 2026 U.S. industry leverage, interest coverage and equity volatility from Damodaran, and official 2027 sector-level probability-of-default adjustments from the NGFS CLIMACRED model. One-year Merton probabilities provide independent baseline credit-risk proxies for 16 non-financial sectors representing \$5.2808 trillion of commitments. Minimax regret is compared with symmetric scenario weighting and maximin under two principal decision scores: expected loss and a credit-compensation payoff using Damodaran synthetic spreads. DAPS raises the commitment-weighted one-year PD from 0.0592% to 0.0897%. At a 12.5% sector cap, minimax regret reduces maximum expected-loss regret by 29.1% relative to symmetric scenario weighting; under the credit-compensation score the reduction is 69.1%. Sensitivity to concentration limits, crosswalk weights, Merton inputs and the climate-PD transport shows that the value of minimax regret is conditional on scenario-specific sector rankings differing near binding portfolio constraints. The contribution is an operational decision-fragility framework for financial institutions and supervisors, not evidence about realized bank behavior or universally optimal climate lending.

We study how to combine diffusion models that disagree about drift and covariance. Candidate-first relative-entropy minimization gives geometric pooling, whereas expert-first minimization gives the arithmetic mixture associated with weighted logarithmic wealth. Different quadratic variations can make path-space entropy infinite, and the arithmetic mixture need not be a Markov diffusion. We therefore specify a local criterion combining drift information, normalized by the second argument's covariance, with quadratic transport between Gaussian shocks in a fixed state metric. A Gaussian identity and an Euler convergence estimate justify this chosen criterion. The expert-first projection has posterior-mean drift and an inverse-covariance penalty for drift dispersion; in one dimension this penalty increases volatility. For Ornstein–Uhlenbeck experts with a common mean-reversion rate, coefficient regularity holds on the full horizon for common volatility and away from the initial time for heterogeneous volatilities. The candidate-first problem has a Hamilton–Jacobi–Bellman characterization. Its matrix covariance selector reduces by congruence to a Bures–Wasserstein barycenter. The condition $H+\lambda M\succ0$, with value Hessian $H$ and state metric $M$, is sharp for finiteness of the unrestricted local covariance problem; compact constraints keep that problem finite. A covariance-disagreement budget interprets the penalty parameter. Linear–quadratic, exact-transition, and financial examples distinguish dynamic volatility reduction, drift-dispersion inflation, and martingale restrictions.