Volatility paths with similar cumulative variance can differ in how that variance accumulates. We study retrospective screening of cumulative forward realized-variance curves using a curvature-informed risk diagnostic (CIRD). The score combines a training-standardized terminal level with a penalty on changes in daily squared-return contributions. This construction provides an interpretable distinction between gradual and spike-dominated volatility accumulation. A split-conformal order statistic calibrates complete block maxima; its finite-sample exceedance bound requires exchangeable block scores. Daily data for four cryptocurrencies over 2019–2024 provide 57 test blocks per asset under a 30-day horizon and seven-day blocks. CIRD’s mean absolute calibration gap is 0.070, compared with 0.130–0.206 for three benchmarks, but this ranking is not statistically established. Recall against a descriptive high–low range reference is 0.631 versus 0.812−0.875. Its 8–10 reference blocks per asset form only two contiguous runs with two distinct peak dates. An uninformative random score also achieves a small gap. Component associations and ablations are consistent with reduced volatility-level sensitivity jointly explaining calibration transfer and lower reference recall. Simulations show substantial finite-sample variation and deteriorating calibration after a volatility-level break. Together, these analyses clarify how curvature weighting affects retrospective screening and calibration transfer across volatility environments. The findings describe a task-specific level–spikiness ordering without establishing detection superiority.
Bitcoin volatility is challenging to forecast because accuracy depends on model choice, information sets, and market regimes. Existing studies often compare models under inconsistent volatility targets, predictors, samples, or evaluation designs, leaving unclear whether forecast gains arise from model architecture, information content, or regime shifts. This study addresses this gap by proposing a unified rolling out-of-sample framework for daily Bitcoin volatility forecasting that compares GARCH-type models, standalone LSTMs, and state-augmented GARCH–LSTM hybrids under the same volatility target, nested information sets, one-step-ahead forecast horizon, and evaluation design, using balanced daily data from 1 February 2018 to 19 July 2025. Predictors are organized into technical (T), technical plus sentiment (TS), and technical, sentiment, and blockchain (TSB) groups to measure the marginal value of added information. Forecast accuracy is assessed mainly with QLIKE, plus RMSE, MAE, and Diebold–Mariano tests across normal and crisis regimes. Predictability is strongly regime-dependent. In normal markets, LSTM-T achieves the lowest loss, with TGARCH and Hybrid_GARCH_T statistically indistinguishable from it. In crises, hybrids minimize QLIKE, but the best specification differs: Hybrid_TGARCH_TS during COVID-19, Hybrid_EGARCH_TSB during the China mining ban, and Hybrid_APARCH_TS during the crypto winter. Hence, hybrid gains are regime-specific and depend on the interaction of volatility dynamics, information sets, and regime characteristics. For daily one-step-ahead Bitcoin forecasts, the evidence supports regime-sensitive modeling and cautions against broad claims of model superiority based on pooled samples.
Stablecoin depegging risk arises when dollar-pegged tokens trade persistently below par. Empirical monitoring requires a defensible definition of when below-peg pressure becomes a depegging event. We develop a two-stage rolling-threshold label from Kraken minute prices: the daily maximum minute-low discount must meet or exceed a coin-specific historical threshold, and at least 120 observed one-minute bars must have lows more than 2.5 basis points below par. The threshold uses the conditional 95th percentile over up to 365 prior observed days. USDT and USDC form the main sample, with DAI examined separately as an extension. Historical-event diagnostics and a fixed 99-cent comparison distinguish relative historical stress from absolute dollar losses. Weighted classifiers are evaluated using walk-forward tests, validation-selected thresholds, common alert budgets, and synchronized calendar-block inference. The operational LightGBM specification achieves precision of 0.233, F1 of 0.230 (95
The digitalization of financial services has generated large volumes of borrower data, enabling the use of advanced analytics in credit risk modeling. This study evaluates the performance of statistical and machine learning models for predicting and updating final Loss Given Default (LGD) estimates during recovery, using proprietary data from Norwegian retail credit card portfolios. Traditional regression models (Ordinary Least Squares, Ridge, and Lasso) are compared with ensemble-based machine learning techniques, including Random Forest and Extreme Gradient Boosting (XGBoost). The results show that machine learning models generally achieve lower prediction error, with XGBoost delivering the strongest predictive performance. However, linear models retain greater interpretability and higher explained variance. The findings highlight trade-offs between predictive accuracy and transparency and provide new empirical evidence on dynamic LGD modeling in unsecured retail credit markets.
Bitcoin’s price follows a well-documented power law over long time scales, yet this trend alone is a poor short-term predictor: at horizons of a few months, today’s price outperforms any trend extrapolation. We construct a deliberately minimal baseline, the Naive–Power Law blend (NvPL), which interpolates smoothly from today’s price (at short horizons) to the fitted power-law trend (at long horizons), and ask whether more elaborate models can beat it out of sample. We compare a broad set of alternatives: mean-reversion and momentum corrections, shrinkage regressions (ridge, LASSO, and elastic net), tree-based machine learning (gradient boosting and random forests), and 18 Bayesian structural time series specifications, some hand-specified and some discovered by an autonomous AI agent (the autoresearch pattern). Every model is evaluated on five non-overlapping two-year periods spanning distinct market conditions (2016–2026). For every primary benchmark candidate—the blend and the shrinkage and tree learners—all tunable settings, including the baseline’s own, are selected on training data alone; the Bayesian specifications are deliberately granted more information as a diagnostic. None reliably improves on NvPL. Formal forecast-comparison tests (Diebold–Mariano, Clark–West, and the Model Confidence Set, with inference that treats the forecast origin as the observational unit) sharpen the picture for the primary candidates: no correction significantly improves on NvPL, and the confidence set retains NvPL while excluding the pure power law. The blend’s structure carries detectable predictive content over a pure no-change forecast (Clark–West), but strictly out of sample this content is too small to reliably separate the two in realized error at one to six months: persistence dominates, and the naive forecast is not excluded from the pooled confidence set. Return-based metrics agree, and no model demonstrates directional accuracy above the always-up rule once forecast dependence is accounted for. Aggressive in-sample tuning, including the agent’s, reverses sharply out of sample, illustrating how greedy search overfits non-stationary financial data. At one to six months, short-term Bitcoin prices are dominated by persistence, and NvPL is a stringent, reproducible benchmark that more complex models should be required to beat.
