Despite the critical role of Loss Given Default (LGD) in regulatory capital calculations under the Basel IRB framework, no prior work has applied a principled Explainable AI (XAI) framework to decompose LGD predictions at the feature level — a gap that limits model governance, regulatory scrutiny, and practitioner insight for unsecured retail lenders. This study addresses that gap by providing what is, to our knowledge, the first application of SHapley Additive exPlanations (SHAP) to LGD estimation. A Light Gradient Boosting Machine (LightGBM) classifier is trained on 9701 defaulted credit card accounts from a Norwegian issuer (September 2023 to September 2025), using a rich pre-default feature set covering transactional, demographic, spending, and macroeconomic variables. TreeSHAP is applied to quantify feature-level predictive importance, and three unsupervised clustering algorithms (K-Means, DBSCAN, Self-Organizing Maps) are evaluated under two dimensionality-reduction methods (PCA and PaCMAP) to test whether behaviourally distinct LGD-based segments exist. The model achieves a test ROC-AUC of 0.814. SHAP-based importance rankings reveal that transactional distress indicators — overdue-to-balance ratio, revolving behaviour, account age, and credit utilisation — dominate prediction, while macroeconomic, demographic, and spending variables contribute negligible incremental value. Removing all transactional features reduces test AUC from 0.814 to 0.669, a decline of 14.5 percentage points. Probability of Default (PD) provides no incremental predictive gain once transactional features are included (account-level Pearson r=0.012). Clustering analysis yields no evidence of stable LGD-based customer segments across any algorithm–embedding combination, suggesting that recovery outcomes vary along a continuous behavioural gradient. The paper makes four contributions: (i) the first SHAP-based decomposition of LGD prediction; (ii) a systematic evaluation of pre-default transactional and spending features as LGD predictors; (iii) an empirical account-level test of PD as an LGD predictor; and (iv) a rigorous multi-algorithm clustering analysis establishing the absence of discrete LGD-based borrower segments.

The spanning hypothesis in government bond markets posits that a small number of yield-curve factors fully summarize the information relevant for pricing and forecasting bond returns. Despite extensive empirical investigation, evidence on the validity of this hypothesis remains mixed. Concurrently, recent advances in machine learning have renewed interest in uncovering nonlinear predictive relationships between macroeconomic conditions and the term structure of interest rates. This paper provides a systematic literature review of studies that employ machine learning methods to forecast bond yields and excess returns, with a particular focus on their implications for the spanning hypothesis. We synthesize evidence across datasets, model classes, validation strategies, and approaches to economic interpretation. The literature is highly fragmented, lacking a coherent model hierarchy or standardized out-of-sample evaluation framework. Nonetheless, evidence from methodologically rigorous studies suggests that nonlinear methods, in particular parsimonious neural networks and tree-based ensembles, can extract predictive signals from macroeconomic data that are not spanned by conventional yield-curve factors. At the same time, we document that reported forecasting gains are frequently overstated due to methodological shortcomings, including inappropriate evaluation metrics, look-ahead bias arising from revised macroeconomic data, and data leakage induced by global pre-processing and model selection. Taken together, our findings imply that while the true predictive advantage of machine learning over traditional term-structure benchmarks is more modest than often claimed, it remains economically meaningful.

This paper examines the literature on applications of artificial intelligence (AI) and machine learning (ML) in assessing credit risk in banks, specifically focusing on probabilistic AI. We uncover unresolved issues and prevalent technical challenges by identifying the predominant AI and ML models employed. Furthermore, we explore potential approaches to address these challenges. The review reveals that research is increasingly exploring AI and ML models for credit risk modeling, often providing more accurate risk predictions than traditional models. Apparently, there is a predominant reliance on traditional deterministic models, highlighting an underutilization of probabilistic models that effectively handle inherent uncertainties in credit risk estimates. Furthermore, we observe that challenges such as overfitting, model interpretability, and regulatory compliance remain prevalent. This suggests significant potential for improving accuracy, reliability, and trustworthiness by increasing applications of probabilistic and generative AI, in combination with explainable AI (XAI).

In the realm of option pricing, parametric models originating from the Black-ScholesMerton framework have proven extremely persistent. However, machine learning models have recently entered the field with success, arguably due to their flexible and non-parametric nature. A combined LSTM-MLP deep learning architecture that combines time series data with cross-sectional pricing information, avoiding explicit volatility estimates, has recently been proposed. This LSTM-MLP model outperforms relevant benchmarks in different dimensions. In this research, we investigated whether a transformer-based alternative is able to better capture the inter-temporal characteristics of the data than the LSTM-based LSTM-MLP model. We found that although the transformer performs better during the extreme market conditions of COVID-19, the LSTM-MLP architecture is overall superior.

This paper presents a deep learning approach for option pricing using a long short-term memory (LSTM) neural network applied to European call options on the S P 500 index. We utilize a rolling window approach that trains 12 instances of the LSTM model, one for each month of 2021. To gain further insight into the model performance, we use explainable artificial intelligence (XAI) through SHapley Additive Explanations (SHAP). We find that the LSTM model outperforms the Black–Scholes and the Heston models and a multilayer perceptron (MLP) neural network regarding overall pricing accuracy. Most notably, the time-sequencing nature of LSTM enables the proposed model to capture sufficient short-term volatility from recently traded options. This result is still robust when controlling for time-varying volatility dynamics. Thus, the model is less prone to measurement errors in volatility.

Traditional option pricing models rely on estimates of expected volatility. The true volatility is not directly observable and must hence be estimated, inevitably with error. Any measurement errors immediately translate into inaccurate pricing, leading to potential losses for economic agents trading options for hedging or speculative purposes. This paper proposes a novel merged LSTM-MLP model for option pricing that circumvents the need for an explicit volatility estimate, leading to more accurate valuations. Through extensive out-of-sample testing on S&P500 call options data from 2015 to 2022 we document the statistical accuracy and economic benefits of the model when compared to relevant benchmarks. The superior performance is enabled by the combined LSTM-MLP architecture, which simultaneously utilizes both time series data and the cross-section of observed option characteristics in a deep learning neural network that accurately captures the complex price dynamics. The results are consistent over time and robust across option moneyness and time-to-expiry.

This paper proposes an explicable early warning machine learning model for predicting financial distress, which generalizes across listed Nordic corporations. We develop a novel dataset, covering the period from Q1 2001 to Q2 2022, in which we combine idiosyncratic quarterly financial statement data, information from financial markets, and indicators of macroeconomic trends. The preferred LightGBM model, whose features are selected by applying explainable artificial intelligence, outperforms the benchmark models by a notable margin across evaluation metrics. We find that features related to liquidity, solvency, and size are highly important indicators of financial health and thus crucial variables for forecasting financial distress. Furthermore, we show that explicitly accounting for seasonality, in combination with entity, market, and macro information, improves model performance.

In this systematic literature review, we examine the existing studies predicting realized volatility and implied volatility indices using artificial intelligence and machine learning. We survey the literature in order to discover whether the proposed methods provide superior forecasts compared to traditional econometric models, how widespread the application of explainable AI is, and to outline potential areas for further research. Generally, we find the efficacy of AI and ML methods for volatility prediction to be highly promising, often providing comparative or better results than their econometric counterparts. Neural networks employing memory, such as Long-Short Term Memory and Gated Recurrent Units, consistently rank among the top performing models. However, traditional econometric models are still highly relevant, commonly yielding similar results as more advanced ML and AI models. In light of the success with ensemble methods, a promising area of research is the use of hybrid models, combining machine learning and econometric models. In spite of the common critique of many machine learning models being of a black-box nature, we find that very few papers apply XAI to analyze and support their empirical results. Thus, we recommend that researchers strive harder to employ XAI in future work. Similarly, we see potential for applications of probabilistic machine learning, effectively quantifying uncertainty in volatility forecasts from machine learning models.

Developing an effective modeling framework to minimize foreign exchange (FX) risk is of vital importance for hedgers and traders in FX markets. In this study, we compare the ability of long short-term memory (LSTM) models to that of random forest and several time series models for forecasting EURUSD implied volatility across the volatility surface. As our literature study argues, there are only a few published papers on this subject. We find that the LSTM model is the best model for shorter option maturities, while the AR-GARCH model is superior when the maturities increase. We observe that the LSTM model is able to capture immense and immediate changes in implied volatility, which is important for hedging against significant shifts in FX rates.

High-frequency tick data have proved helpful for forecasting volatility across asset classes. In the finite samples typically faced by practitioners, however, noise inherent in tick-level prices creates inaccuracies in model parameter estimates and resulting forecasts. A remedy proposed to alleviate these measurement errors is to include higher-order moments, more specifically the realized quarticity, in volatility prediction models. In this paper, we investigate the relevance of this approach in foreign exchange markets, as represented by EURUSD and USDJPY data from 2010 to 2022. Using well-established realized volatility models, we find that including realized quarticity leads to higher precision in daily, weekly, and monthly out-of-sample forecasts. These results are robust across estimation windows, evaluation metrics, and model specifications.

Global energy production is undergoing a transition from fossils to renewables. At the same time, the Norwegian Oil Fund has grown exponentially in size and is now a major global investor. These events in combination are likely to impact the dynamics of the Norwegian krone. Concurrently, the persistent weakening of the Norwegian krone (NOK), hitting record low exchange rates against the major currencies, is sparking national and international interest. Using updated data, we find that oil prices and global asset prices are both important drivers of EURNOK returns. However, we find that the relative importance changed following the 2015 oil price decline, whereafter asset prices became more significant. Furthermore, we observe an impact of investor risk aversion, suggesting that the krone is no longer a safe-haven currency.

In this study, we propose a semiparametric, parsimonious value-at-risk forecasting model, based on quantile regression and machine learning methods, combined with readily available market prices of option contracts from the over-the-counter foreign exchange rate interbank market. We aim at improving existing methods for VaR prediction of currency investments using machine learning. We employ two different methods, i.e., ensemble methods and neural networks. Explanatory variables are implied volatilities with plausible economic interpretation. The forward-looking nature of the model, achieved by the application of implied volatilities as risk factors, ensures that new information is rapidly reflected in value-at-risk estimates. To the best of our knowledge, this study is the first to utilize information in the volatility surface, combined with machine learning and quantile regression, for VaR prediction of currency investments. The proposed ensemble models achieve good estimates across all quantiles. The light gradient boosting machine model and the categorical boosting model both yield estimates which are better than, or equal to, those of the benchmark model. In general, neural network models are quite unstable.

Fundamentally, the term premium in long-term nominal yields is compensation to investors for bearing interest rate risk. There is substantial evidence of sizable and time-varying term premia. As opposed to yields, term premia are not directly observable. In this paper, we estimate term premia in Norwegian interest rate swaps from a set of dynamic term structure models, covering the period from 2001/04 until 2022/06. In line with international studies, we find evidence of declining term premia over the sample period. Furthermore, our estimates indicate that term premia have been close to zero, as well as negative in periods, during the last decade of global extraordinary monetary policy measures. We find that the recent rise in Norwegian interest rate swaps is partly caused by increases in term premia. From a practitioner’s perspective, our term premia estimates can be utilized as part of applied management of both investment and debt portfolios.

In this paper we propose a semi-parametric, parsimonious value-at-risk forecasting model based on quantile regression and readily available market prices of option contracts from the over-the-counter foreign exchange interbank market. Explanatory variables are implied volatilities with plausible economic interpretation. The forward-looking nature of the model, induced by the application of implied moments as risk factors, ensures that new information is rapidly reflected in value-at-risk estimates. The proposed model outperforms traditional benchmark models when evaluated in-sample and out-of-sample on EUR/USD data. The model is relatively easy to estimate, which facilitates practical application. Our quantile regression implied moments model is subjected to extensive risk model validation by means of backtesting, using both coverage tests and loss functions. Thus, his paper is relevant for both risk modeling and risk model validation in the context of foreign exchange risk.

Principal component analysis (PCA) is well established as a powerful statistical technique in the realm of yield curve modeling. PCA based term structure models typically provide accurate fit to observed yields and explain most of the cross-sectional variation of yields. Although principal components are building blocks of modern term structure models, the approach has been less explored for the purpose of risk modelling—such as Value-at-Risk and Expected Shortfall. Interest rate risk models are generally challenging to specify and estimate, due to the regime switching behavior of yields and yield volatilities. In this paper, we contribute to the literature by combining estimates of conditional principal component volatilities in a quantile regression (QREG) framework to infer distributional yield estimates. The proposed PCA-QREG model offers predictions that are of high accuracy for most maturities while retaining simplicity in application and interpretability.