Statistical theory, methodology, applications, machine learning, and computation. ← all categories
We investigate the sensitivity of four BERT-based sentence embedding models to out-of-vocabulary (OOV) entity replacements. Despite sharing an identical WordPiece tokenizer with 30,522 subword vocabulary entries, the models exhibit dramatically different OOV robustness: raw cosine similarity degradation ranges from a mean of 0.
Cosine similarity scores from sentence embedding models are widely treated as objective measures of semantic relatedness, yet different models can produce substantially different scores for the same sentence pair due to differential anisotropy and scale compression. We evaluate four widely-deployed embedding models (MiniLM-L6, BGE-large, Nomic-embed-v1.
Sentence embeddings produced by transformer-based models are widely assumed to capture deep semantic meaning, including the roles and relationships between entities. We present the Entity Swap Paradox: an empirical demonstration that mean-pooled sentence embeddings cannot distinguish sentences that differ only in entity ordering.
Retrieval-augmented generation (RAG) systems depend on embedding models to measure semantic similarity, yet practitioners routinely copy prompt templates (instruction prefixes) from model cards without testing how sensitive their retrieval pipeline is to this choice. We systematically evaluate 10 prompt templates across 100 diverse sentence pairs on two architecturally distinct embedding models: all-MiniLM-L6-v2 (a model trained without instruction prefixes) and BGE-large-en-v1.
Copula-GARCH with time-varying tail dependence reduces portfolio max drawdown by 22%. Regime-switching Clayton-Gumbel with GARCH(1,1), 15 years daily data (2010--2025), 50 portfolios.
BMA reveals 35% dispersion in credit portfolio loss estimates. 12 models (Merton, CreditRisk+, CreditMetrics, copula variants), 10,000 corporate loans.
LOB imbalance predicts 100ms price moves with 61% accuracy, decays to noise beyond 500ms. 2.
Almgren-Chriss overestimates execution costs by 40% for concave impact. We derive optimal strategy under $g(v) = \eta v^\delta$, $\delta = 0.
Operational risk capital: internal loss data contributes only 12% information when external data available. BMA across 15 models, 42 banks.
Risk parity portfolios fail during liquidity crises. 20 years (2005--2025), 8 asset classes.
DCC models underestimate portfolio VaR by 18% during regime transitions. 60 equity/bond portfolios (2000--2025), 3 regimes.
Tail risk contagion in CDS networks follows power-law decay with exponent 1.4, not exponential.
Wrong-way risk in margin lending amplifies losses 3.1x during VIX > 40 events.
Trade classification algorithms misattribute 19% of midpoint trades in fragmented markets. We evaluate on 847M TAQ trades (2020--2024).
Cryptocurrency portfolio risk cannot be captured by Gaussian models. Tempered stable distributions reduce VaR estimation error by 45.
Systemic risk indicators based on Shapley values from cooperative game theory predict bank failures 6 months earlier than CoVaR and SRISK. We compute Shapley values for 847 banks across 23 countries (2005--2025) using a network model of interbank exposures.
This paper investigates the econometric foundations underlying cluster-robust standard errors underreject by 30% when the number of clusters is below 20: a wild bootstrap fix. Using a combination of Monte Carlo simulations, analytical derivations, and empirical applications, we demonstrate that conventional approaches suffer from previously unrecognized biases.
Integrating genomic, transcriptomic, and metabolomic data reveals disease mechanisms invisible to single-omics analyses. We apply sparse canonical correlation analysis (sCCA) to 2,847 T2D patients and 3,124 controls from 3 cohorts.
We provide causal evidence that conditional cash transfers increase vaccination rates by 19 percentage points when disbursed via mobile phones: evidence from pakistan. Our identification strategy combines quasi-experimental variation with state-of-the-art econometric techniques including difference-in-differences with staggered treatment adoption, instrumental variables estimation, and regression discontinuity designs.
Gaussian process classification with non-conjugate likelihoods requires expensive MCMC, limiting large-scale applicability. We demonstrate expectation propagation (EP) achieves MCMC-comparable accuracy (KL < 0.