We devise an innovative technical analysis strategy that leverages combinations of technical indicators and machine learning techniques. Our models use image representations generated by visualizing the time-series data of individual stocks over a specific period to identify price patterns. To avoid data snooping, we carefully curtail the size of our training sets and shuffle the sample sets between models. Our approach yields highly reliable and impactful predictions on the U.S. stock market from January 1992 to December 2022 compared to multiple benchmarks, including momentum and reversal, traditional technical trading rules, industrial portfolios, and the Fama–French five-factor model. Using the GradCAM method, we visualize the attention maps of model predictions, providing strong interpretability to our approach of technical analysis.
We propose an easy-to-implement framework for combining quantile forecasts, applied to forecasting GDP growth. Using quantile regressions, our combination scheme assigns weights to individual forecasts from different indicators based on quantile scores. Previous studies suggest distributional variation in forecasting performance of leading indicators: some indicators predict the mean well, while others excel at predicting the tails. Our approach leverages this by assigning different combination weights to various quantiles of the predictive distribution. In an empirical application to forecast US GDP growth using common predictors, forecasts from our quantile combination outperform those from commonly used combination approaches.
We apply functional data analysis to survey expectations data, and show that functional principal component analysis combined with functional regression analysis is a fruitful way of capturing the effects of others’ forecasts on a respondent’s inflation forecasts. We estimate forward-looking Phillips curves on each respondent’s inflation and unemployment rate forecasts, and show that for nearly half of the respondents, the forecasts of others are important. The functional principal components of the cross-sectional distributions of forecasts are shown to capture characteristics other than the mean or consensus forecast, and include forecaster disagreement.
In recent years, survival models have received increasing attention in credit risk. Unlike classification models typically used to model defaults, survival analysis can model not only whether a borrower will default but also the time to default. Since loan transaction data are discrete time-series information, it is natural to apply a discrete-time survival model (DTSM). At the same time, there have been significant advances in deep neural networks. In this paper, we extend the DTSM using long short-term memory (LSTM) networks, incorporating an adapted LSTM-based attention mechanism to better uncover temporal features that influence the probability of default, along with a washout phase, which is used to iterate several dummy LSTM prediction steps and hence addresses the LSTM state initialization problem. Finally, instead of using a standard attention mechanism that linearly encodes the input, we adapt it with an LSTM layer that captures non-linear temporal features from the sequential data. This model shows great improvement in model fit and predictive ability, in comparison with baseline linear DTSMs and standard LSTM with attention, when evaluated on US mortgage data. Moreover, we show that our proposed model gives good forecast performance, providing practitioners with a practical and powerful early warning service to manage potential credit loss.
We propose a novel machine learning approach for forecasting the distribution of stock returns using a rich set of firm-level and market predictors. Our method combines a two-stage quantile neural network with spline interpolation to construct smooth, flexible cumulative distribution functions without relying on restrictive parametric assumptions. This allows for accurate modelling of non-Gaussian features such as fat tails and asymmetries. Furthermore, we show how to derive other statistics from the forecasted return distribution, such as the mean, variance, skewness, and kurtosis. The derived mean and variance forecasts offer significantly improved out-of-sample performance compared to standard models. We demonstrate the robustness of the method in U.S. and international markets.
Effective early identification of potentially fragile financial institutions is crucial for mitigating systemic risks due to the strong interconnection within the financial sector. For the first time, we apply the Structural Artificial Neural Network (SANN) to predict failure of financial institutions, using insurance companies as the context. The SANN is a category-wise machine learning algorithm (i.e., a category-aware learner) that extracts nonredundant information from Big Data with different categories of information. It enables the grouping of predictors into economic categories and thus allows interactions both within and across categories. We show that the SANN significantly improves the out-of-sample predictions of insurer failures, resulting in an average increase in the area under the curve (AUC) of 1.83% to 9.55% compared to traditional machine learning and logistic models, based on a sample of 2424 insurers from 17 European countries. We also document that macroeconomic and yield information provides nonredundant information in forecasting the failures, in addition to firm characteristics.
Classical field forecast evaluations mainly rely on local scores such as the RMSE or MAE. These metrics severely over-penalize small spatial or temporal displacements of coherent structures, a limitation known as the double-penalty issue and common to many forecasting domains. The present paper introduces a tolerance-based framework built on the three-dimensional γ-index, initially designed for medical dose verification, as a unified acceptance criterion for gridded forecasts. The method embeds explicit margins in space (DTA), time (TTA), and intensity (IDT) and evaluates whether predictions agree with observations within predefined physical bounds rather than through pixel-wise differences only. A synthetic illustration is first used to show why conventional metrics can misrepresent usable forecasts. The approach is then applied to satellite-derived surface solar irradiance fields to demonstrate operational behavior on a real dataset. The results confirm that the γ-criterion preserves structural consistency under minor positional noise while isolating physically significant discrepancies. The formulation is generic and can be implemented for any gridded variable, provided that meaningful tolerances are defined. It offers a pragmatic complement to existing spatial verification tools in general forecasting workflows.
Effective energy trading requires probabilistic forecasts to quantify uncertainty and manage financial risk. In this paper, we describe our approach, which combines separate wind and solar models using a state-of-the-art stacked CatBoost framework. The effectiveness of this method was validated in the HEFTCom2024 competition, where it was the winning entry for forecasting and trading the combined generation of a 1200 MW offshore wind and 2400 MW solar portfolio in England. Key factors contributing to our success include the use of separate wind and solar power models, the incorporation of three different weather forecast datasets, no missed submissions (benchmark fills), and effective handling of a long-lasting cable issue for the offshore wind farm. Although the main focus was on the forecasting model, we also won the trading track. This is attributed mainly to our forecast accuracy, but our trading score exceeded expectations based on the trend in trading vs. the forecasting scores of co-competitors. (c) 2026 The Authors. Published by Elsevier B.V. on behalf of International Institute of Forecasters. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Financial distress prediction (FDP) is paramount for economic stability in an increasingly volatile global landscape. Existing statistical and machine learning models confront two persistent challenges: the inherent class imbalance in financial datasets, where distress cases are rare; and the poor interpretability of complex black-box models, which hinders trust in high-stakes contexts. To address these limitations, we propose the XGBoost-CSMA, a reweighting-based boosting ensemble framework for interpretable imbalanced financial distress prediction. Departing from cost-sensitive methods that rely on static, expert-defined cost matrices, our confidence-scaled margin adaptation (CSMA) objective dynamically adjusts predictive margins based on confidence levels. This enables automatic adaptation to varying financial distress signals without predefined costs, mitigating imbalance while preserving discriminative power for critical borderline cases. Furthermore, XGBoost-CSMA embeds TreeSHAP explanations as an intrinsic component of model validation, creating a feedback loop that enhances both predictive performance and decision transparency. We also provide the theoretical foundations for efficiently integrating cost-sensitive learning into gradient boosting. Empirical assessments utilizing real-world corporate financial datasets substantiate the superior efficacy of XGBoost-CSMA relative to state-of-the-art baselines, a conclusion reinforced by substantial improvements in the geometric mean and true positive rate that underscore its enhanced sensitivity to distress signals. Furthermore, this predictive advancement is coupled with the provision of transparent risk interpretations, thereby ensuring that the model delivers both high diagnostic accuracy and actionable decision support. (c) 2026 International Institute of Forecasters. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Assessing the risks of having either very low or very high inflation is crucial for policy-makers, businesses, and householders. In a globalised world, these risks are increasingly determined by international conditions. In this paper, we empirically analyse the impact of international inflation factors on forecasting monthly domestic inflation risks in a large number of economies observed worldwide from 1999 to 2022. Risk forecasts are obtained using factor-augmented quantile regressions estimated with international factors extracted from a multi-level dynamic factor model with overlapping blocks of inflations corresponding to economies grouped either in a given geographical region or according to their development level. We conclude that in a large number of countries, international factors are relevant to explain the right tail of the distribution of inflation, and consequently they are more relevant for the risk related to high inflation than for low inflation. The role of international factors is stronger in developed European countries, while the inflation risks of low-income developing countries are hardly affected by international conditions, and the results for middle-income countries are mixed. We also show that the predictive power of international factors has increased in the most recent years of high inflation. (c) 2025 The Author(s). Published by Elsevier B.V. on behalf of International Institute of Forecasters. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
The rise of omnichannel grocery retail has introduced significant operational complexities, emphasizing the importance of high-frequency demand forecasting. In this context, achieving both accuracy and computational efficiency with standard forecasting tools remains a challenge. Addressing these limitations, we propose a novel framework that integrates deep learning models with a decoupled approach that separates structural demand modeling from short-term fluctuation prediction to enhance prediction accuracy. Our method combines neural hierarchical interpolation for time series forecasting (N-HiTS) and a mixture density network (MDN) to capture short-term fluctuations and structural demand patterns, respectively. This framework is extended to probabilistic forecasting, comparing quantile-based and distributional models, both with and without the decoupling approach. Empirical validation using data from a leading on-demand delivery service demonstrates significant improvements in deep learning methods over traditional ARIMA methods and the industry-standard gradient boosting machine (GBM), and validates the effectiveness of the decoupling approach. The new framework reduces the mean absolute percentage error (MAPE) for point estimates from 23.00% to 14.31% (a 37.78% reduction) and the continuous ranked probability score (CRPS) from 10.85 to 2.34 (a 78.44% reduction). These findings can provide grocery e-commerce companies with valuable insights for optimizing inventory management, driver scheduling, and overall operational efficiency.
The CoVaR and CoES are two of the most widely used measures of systemic risk in economics and finance. In this paper, we introduce a novel quantile regression approach for jointly estimating CoVaR and CoES. This method extends existing Asymmetric Laplace (AL) joint estimation techniques for Value at Risk (VaR) and Expected Shortfall (ES) to the realm of systemic risk measures. We generalize the joint quantile regression model to a time-varying setting by allowing the multiplicative factor between CoVaR and CoES to vary over time using a score-driven dynamic approach. We address the inference problem by developing a suitable likelihood-based Expectation-Maximization (EM) algorithm. We apply the new model to real data from the Chinese stock market, covering the period from 2010 to 2023. The results indicate that our models outperform their constant multiplicative factor counterparts and alternative GARCH-type models. In risk management applications, we construct a Skewness Mean-Variance (SMV) portfolio to manage risk exposure to system-wide distress. Finally, we employ network techniques to capture systemic risk contagion among industries and to monitor the market's systemic risk level dynamically. (c) 2026 International Institute of Forecasters. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
The availability of rich online data has opened new opportunities for election forecasting. While typical election forecasting predicts results at the national level, the accumulation of information on candidate and voter behavior enables making predictions on a more granular level. Most studies using online data focus on contests with a small number of candidates, leaving a research gap for elections with larger candidate pools. Elections with numerous candidates differ from races with a limited number of candidates, as voters are more inclined to use heuristics and mental shortcuts when selecting their preferred candidate. Building on this insight, this paper introduces a model to predict each candidate's vote share in the context of Finnish parliamentary elections. An ex ante forecast based on the model was published before the 2023 Finnish parliamentary election, which correctly identified 150 of the 200 candidates elected to parliament from a total pool of 2468 contestants. The results showcase the potential to effectively leverage the rich online data environment, thus complementing existing methodologies. Compared to traditional approaches, the proposed model provides candidate-level estimates, which offer insights into intra-party competition and list rankings. (c) 2026 The Author. Published by Elsevier B.V. on behalf of International Institute of Forecasters. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
In this paper, we present a novel approach for probabilistic forecasting based on expectile smoothing of river network data. The Miho River dataset, which is the focus of this study, contains spatio-temporal observations across a stream network. As the inherent structure of the stream network should be taken into account and time points are irregular and vary across observation sites, developing a forecasting method presents significant challenges. To address this, we extend the flexible smoothing method of spatio-temporal river stream network data by incorporating expectile regression. We represent expectile curves of annual observations as a function of time and employ the forecasting method of functional time series. Through expectile regression, we extract information beyond the mean response for river network data analysis and develop a probabilistic forecasting method by predicting the expectile process. We demonstrate the results of the proposed method with the Miho River data and evaluate its performance.
Theoretical research suggests potential cross-currency predictability in currency exchange markets, but empirical findings with macro-based models show mixed results. This has led to a growing focus on micro-based models which explore how trading integrates fundamental information into exchange rates. This study stands out by using detailed limit order book data for the foreign exchange market, enabling microlevel insights. It covers multiple currency pairs, assessing cross-currency predictability. Emphasizing short-term forecasts from one minute to one hour, it introduces factoraugmented regressions, including unsupervised and supervised principal components analysis to tackle data dimensionality, and contrasts those to LASSO and random forest methods. Our findings reveal generally low predictability across various models, supporting the efficient market hypothesis. We find that cross-currency variables offer limited additional insight, but certain microstructure variables like order flow show short-term predictive power, suggesting transient market inefficiencies. (c) 2026 The Authors. Published by Elsevier B.V. on behalf of International Institute of Forecasters. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
