Stefan Lessmann
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Stefan Lessmann has a total of 134 co-authors. Here are the top co-authors by number of shared publications:
- Stefan Voß
Institute of Information Systems, University of Hamburg, 20146 Hamburg, Germany
7 shared publications - Daniel Traian Pele
Bucharest University of Economic Studies , Bucharest , Romania
6 shared publications - Wolfgang Karl Härdle
IDA Institute for Digital Assets, Bucharest University of Economic Studies, Bucharest, Romania
5 shared publications Ming-Chien Sung
Centre for Risk Research, School of Management, University of Southampton, Highfield, Southampton SO17 1BJ, United Kingdom
4 shared publications- Alona Zharova
Humboldt-Universität zu Berlin
3 shared publications - Bart Baesens
KU Leuven
3 shared publications
Latest Publications
- Open Access
- 10 September 2026
- Published by Elsevier in SSRN Electronic Journal
Examining the Impact of Bias Mitigation on High-stakes Domains: Balancing Performance, Fairness, and Explainability
Machine Learning models are increasingly used in high-stakes decision-making, necessitating that algorithmic decisioning systems are not only accurate but also explainable and guarantee fair treatment of affected subjects. Counterfactual explanations have emerged as a promising approach to address these requirements. They offer intuitive, user-centered justifications by illustrating how small input changes can lead to different model outcomes. Likewise, bias mitigation techniques improve group fairness and help financial institutions ensure compliance in lending operations. To that end, bias mitigation techniques alter a classifier's training data, the classifier itself, or the probability predictions it generates. Impacting the prediction model development process and potentially the underlying data, bias mitigation may substantially impact explanation methods, particularly counterfactual explanations. However, the effect of bias mitigation on explainability remains unclear. Moreover, high-stakes Machine Learning systems must maintain high predictive performance, which some bias mitigation techniques may compromise. Given the complex interplay between accuracy, explainability, and fairness, this paper examines their dependency structure. We aim to clarify interactions and synergies among learning algorithms, bias mitigation techniques, and explanation methods to help decision-makers navigate the comprehensive option space and ultimately design powerful and compliant systems. Our empirical analysis embraces multiple bias mitigation techniques, spanning pre-, in-, and post-processing methods, established learning algorithms, and multiple real-world data sets with diverse characteristics. Unsurprisingly, no single bias mitigation method excels across all evaluation dimensions. Some methods improve fairness at the cost of accuracy or explanation quality, while others benefit predictive performance or explanation quality but may compromise model fairness. Overall, our results highlight the need for joint evaluation frameworks that consider predictive performance, fairness, and explainability as interdependent components in trustworthy AI systems.
- 10 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessExamining the Impact of Bias Mitigation on High-stakes Domains: Balancing Performance, Fairness, and Explainability
Machine Learning models are increasingly used in high-stakes decision-making, necessitating that algorithmic decisioning systems are not only accurate but also explainable and guarantee fair treatment of affected subjects. Counterfactual explanations have emerged as a promising approach to address these requirements. They offer intuitive, user-centered justifications by illustrating how small input changes can lead to different model outcomes. Likewise, bias mitigation techniques improve group fairness and help financial institutions ensure compliance in lending operations. To that end, bias mitigation techniques alter a classifier's training data, the classifier itself, or the probability predictions it generates. Impacting the prediction model development process and potentially the underlying data, bias mitigation may substantially impact explanation methods, particularly counterfactual explanations. However, the effect of bias mitigation on explainability remains unclear. Moreover, high-stakes Machine Learning systems must maintain high predictive performance, which some bias mitigation techniques may compromise. Given the complex interplay between accuracy, explainability, and fairness, this paper examines their dependency structure. We aim to clarify interactions and synergies among learning algorithms, bias mitigation techniques, and explanation methods to help decision-makers navigate the comprehensive option space and ultimately design powerful and compliant systems. Our empirical analysis embraces multiple bias mitigation techniques, spanning pre-, in-, and post-processing methods, established learning algorithms, and multiple real-world data sets with diverse characteristics. Unsurprisingly, no single bias mitigation method excels across all evaluation dimensions. Some methods improve fairness at the cost of accuracy or explanation quality, while others benefit predictive performance or explanation quality but may compromise model fairness. Overall, our results highlight the need for joint evaluation frameworks that consider predictive performance, fairness, and explainability as interdependent components in trustworthy AI systems.
- Open Access
- 2 September 2026
- Published by Springer Nature in Management & Marketing
- Vol. 21 (1) , 21
Toward scalable generative AI: efficient language model distillation via zero-shot rationales
This paper investigates an efficient approach for distilling Large Language Models (LLMs) into smaller, application-specific models using zero-shot Chain of Thought (CoT) rationale generation and Optimization by Prompting (OPRO). To address the challenges of deploying computationally intensive generative AI for narrow tasks or resource-constrained environments, the approach leverages LLM reasoning capabilities to generate both labels and natural language explanations for unlabeled data. By reducing reliance on human-generated annotations, the approach substantially lowers annotation requirements and prompting costs while maintaining comparable performance in the evaluated settings. We formulate distillation as a multi-task learning problem in which student models are trained to jointly predict labels and learn from teacher-generated rationales, with the goal of improving data efficiency and generalization. Building on established zero-shot Chain of Thought (CoT) prompting and the OPRO prompt optimization technique, we use teacher-generated rationales to reduce annotation token requirements and examine the associated performance and efficiency gains. Additionally, we systematically investigate how explanation properties affect distillation efficiency. Across natural language inference and question answering benchmarks, results indicate that near-optimal performance can be achieved even when rationales are provided for only a subset of the training data, and that shorter explanations are often sufficient. These findings provide practical insights into the trade-offs between rationale generation cost and student model performance. Overall, this work contributes empirical evidence on the effectiveness and cost characteristics of rationale-based distillation for training compact, task-specific language models with minimal human intervention.
- 2 September 2026
- Published by Springer Nature in Management & Marketing
- Vol. 21 (1) , 21
Open AccessToward scalable generative AI: efficient language model distillation via zero-shot rationales
This paper investigates an efficient approach for distilling Large Language Models (LLMs) into smaller, application-specific models using zero-shot Chain of Thought (CoT) rationale generation and Optimization by Prompting (OPRO). To address the challenges of deploying computationally intensive generative AI for narrow tasks or resource-constrained environments, the approach leverages LLM reasoning capabilities to generate both labels and natural language explanations for unlabeled data. By reducing reliance on human-generated annotations, the approach substantially lowers annotation requirements and prompting costs while maintaining comparable performance in the evaluated settings. We formulate distillation as a multi-task learning problem in which student models are trained to jointly predict labels and learn from teacher-generated rationales, with the goal of improving data efficiency and generalization. Building on established zero-shot Chain of Thought (CoT) prompting and the OPRO prompt optimization technique, we use teacher-generated rationales to reduce annotation token requirements and examine the associated performance and efficiency gains. Additionally, we systematically investigate how explanation properties affect distillation efficiency. Across natural language inference and question answering benchmarks, results indicate that near-optimal performance can be achieved even when rationales are provided for only a subset of the training data, and that shorter explanations are often sufficient. These findings provide practical insights into the trade-offs between rationale generation cost and student model performance. Overall, this work contributes empirical evidence on the effectiveness and cost characteristics of rationale-based distillation for training compact, task-specific language models with minimal human intervention.
- Open Access
- 17 August 2026
- Published by Taylor & Francis in Journal of the Operational Research Society
- p. 1-15
Incorporating data drift to perform survival analysis on credit risk
- 17 August 2026
- Published by Taylor & Francis in Journal of the Operational Research Society
- p. 1-15
Open AccessIncorporating data drift to perform survival analysis on credit risk
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Energy Financial Risk Meter: Systemic Risk in the EU Energy Sector
This paper introduces the Energy Financial Risk Meter (EFRM), a sector-specific systemic risk indicator for the European Union's energy sector. The EFRM is constructed by applying LASSO quantile regression to daily stock returns and market capitalisation data from 100 major European energy companies, while also integrating a set of economic variables and a geopolitical risk index as exogenous drivers. The EFRM index captures major systemic events in the energy markets, namely the 2020 COVID pandemic and the 2022 Russia-Ukraine war. By employing rolling-window VAR and time-varying parameter VAR (TVP-VAR) estimations across monthly and daily frequencies, we document the activation of a previously dormant geopolitical-to-energy-risk transmission channel and the subsequent material attenuation of downstream transmission to European equity and credit markets. Cross-continental validation using US Treasury yields and the US Energy SPDR ETF confirms the global external validity of the EFRM signal. From a practical perspective, the proposed indicator offers policymakers, regulators, and financial stability authorities a high-frequency tool for early detection of systemic threats in one of the most critical components of the EU economy.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEnergy Financial Risk Meter: Systemic Risk in the EU Energy Sector
This paper introduces the Energy Financial Risk Meter (EFRM), a sector-specific systemic risk indicator for the European Union's energy sector. The EFRM is constructed by applying LASSO quantile regression to daily stock returns and market capitalisation data from 100 major European energy companies, while also integrating a set of economic variables and a geopolitical risk index as exogenous drivers. The EFRM index captures major systemic events in the energy markets, namely the 2020 COVID pandemic and the 2022 Russia-Ukraine war. By employing rolling-window VAR and time-varying parameter VAR (TVP-VAR) estimations across monthly and daily frequencies, we document the activation of a previously dormant geopolitical-to-energy-risk transmission channel and the subsequent material attenuation of downstream transmission to European equity and credit markets. Cross-continental validation using US Treasury yields and the US Energy SPDR ETF confirms the global external validity of the EFRM signal. From a practical perspective, the proposed indicator offers policymakers, regulators, and financial stability authorities a high-frequency tool for early detection of systemic threats in one of the most critical components of the EU economy.
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Mapping Causal Machine Learning in Energy Finance Domain
The rapid diffusion of machine learning techniques in financial and energy market research has coincided with a renewed emphasis on causal inference frameworks that deliver structurally interpretable and policy-relevant insights. Although these domains have developed extensively in parallel, their systematic integration remains limited and insufficiently mapped. This paper surveys the literature on the intersection of causal inference, energy finance, and machine learning. First, we conduct a bibliometric analysis using a systematic search strategy organized into three conceptual blocks: causal inference, energy finance, and machine learning. Second, we identify and synthesize the principal methodological commonalities across the selected studies. We classify contributions by their empirical settings and data structures, highlighting how causality is operationalized in data-driven energy finance applications. Third, we provide an indepth methodological overview of the main approaches employed in this literature.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMapping Causal Machine Learning in Energy Finance Domain
The rapid diffusion of machine learning techniques in financial and energy market research has coincided with a renewed emphasis on causal inference frameworks that deliver structurally interpretable and policy-relevant insights. Although these domains have developed extensively in parallel, their systematic integration remains limited and insufficiently mapped. This paper surveys the literature on the intersection of causal inference, energy finance, and machine learning. First, we conduct a bibliometric analysis using a systematic search strategy organized into three conceptual blocks: causal inference, energy finance, and machine learning. Second, we identify and synthesize the principal methodological commonalities across the selected studies. We classify contributions by their empirical settings and data structures, highlighting how causality is operationalized in data-driven energy finance applications. Third, we provide an indepth methodological overview of the main approaches employed in this literature.
- Open Access
- 13 July 2026
- Published by Elsevier in SSRN Electronic Journal
Event-Based Limit Order Book Representations for Probabilistic VWAP Forecasting in Intraday Electricity Markets
The volatile nature of renewable energy sources lead to a large set of regimes in the intraday continuous electricity market. In order to make operational decisions in the market, existing forecasting approaches often use fixed clock-time bar construction and sample observations in equidistant time steps. Thus, clock-based sampling treats different levels of activity equally and ignores the volume or contracts traded and thus may distort inference and reduce the usefulness of forecasts when accurate uncertainty quantification is of particular interest. In contrast to the frequently used clock-based bar construction method, this paper considers how activity-aligned representations of high-frequency data affect probabilistic forecasting performance. Therefore, it uses event-based bar construction schemes based on trade counts and volumes of the limit order book and transaction data. The study places particular focus on traded-value bars to align comparable levels of executed market exposure. As a meaningful benchmark for market consensus and execution quality, the forecasting target is chosen to be the volume-weighted average price (VWAP). The paper provides, to the best of our knowledge, the first systematic comparison of event-based limit order book bar construction schemes in continuous intraday electricity markets. The results show that the choice of data representation is a first-order determinant of probabilistic forecasting performance. Further, event-based bar construction consistently exceeds clock-time bar construction in forecast accuracy and calibration. Especially traded-value bars deliver strong forecasting performance. The operational value of event-based sampling is demonstrated in two real-world decision settings and tested for performance in regime diagnostics and portfolio balancing. In regime diagnostic applications, event-based representations yield more reliable, better-calibrated market-state probabilities. In portfolio-balancing applications, traded-value bars improve risk-adjusted performance. They also provide more robust scale-normalised balancing quality across execution thresholds. Overall, the findings show that the representation of irregular market activity is central to modelling continuous intraday electricity markets and improved data representations can support downstream applications such as balancing, risk management, and flexible asset optimisation.
- 13 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEvent-Based Limit Order Book Representations for Probabilistic VWAP Forecasting in Intraday Electricity Markets
The volatile nature of renewable energy sources lead to a large set of regimes in the intraday continuous electricity market. In order to make operational decisions in the market, existing forecasting approaches often use fixed clock-time bar construction and sample observations in equidistant time steps. Thus, clock-based sampling treats different levels of activity equally and ignores the volume or contracts traded and thus may distort inference and reduce the usefulness of forecasts when accurate uncertainty quantification is of particular interest. In contrast to the frequently used clock-based bar construction method, this paper considers how activity-aligned representations of high-frequency data affect probabilistic forecasting performance. Therefore, it uses event-based bar construction schemes based on trade counts and volumes of the limit order book and transaction data. The study places particular focus on traded-value bars to align comparable levels of executed market exposure. As a meaningful benchmark for market consensus and execution quality, the forecasting target is chosen to be the volume-weighted average price (VWAP). The paper provides, to the best of our knowledge, the first systematic comparison of event-based limit order book bar construction schemes in continuous intraday electricity markets. The results show that the choice of data representation is a first-order determinant of probabilistic forecasting performance. Further, event-based bar construction consistently exceeds clock-time bar construction in forecast accuracy and calibration. Especially traded-value bars deliver strong forecasting performance. The operational value of event-based sampling is demonstrated in two real-world decision settings and tested for performance in regime diagnostics and portfolio balancing. In regime diagnostic applications, event-based representations yield more reliable, better-calibrated market-state probabilities. In portfolio-balancing applications, traded-value bars improve risk-adjusted performance. They also provide more robust scale-normalised balancing quality across execution thresholds. Overall, the findings show that the representation of irregular market activity is central to modelling continuous intraday electricity markets and improved data representations can support downstream applications such as balancing, risk management, and flexible asset optimisation.
- Open Access
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
Energy Financial Risk Meter: Systemic Risk in the EU Energy Sector
This paper introduces the Energy Financial Risk Meter (EFRM), a sector-specific systemic risk indicator for the European Union's energy sector. The EFRM is constructed by applying LASSO quantile regression to daily stock returns and market capitalisation data from 100 major European energy companies, while also integrating a set of economic variables and a geopolitical risk index as exogenous drivers. The EFRM index captures major systemic events in the energy markets, namely the 2020 COVID pandemic and the 2022 Russia--Ukraine war. By employing rolling-window VAR and time-varying parameter VAR (TVP-VAR) estimations across monthly and daily frequencies, we document the activation of a previously dormant geopolitical-to-energy-risk transmission channel and the subsequent material attenuation of downstream transmission to European equity and credit markets. Cross-continental validation using US Treasury yields and the US Energy SPDR ETF confirms the global external validity of the EFRM signal. From a practical perspective, the proposed indicator offers policymakers, regulators, and financial stability authorities a high-frequency tool for early detection of systemic threats in one of the most critical components of the EU economy.
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEnergy Financial Risk Meter: Systemic Risk in the EU Energy Sector
This paper introduces the Energy Financial Risk Meter (EFRM), a sector-specific systemic risk indicator for the European Union's energy sector. The EFRM is constructed by applying LASSO quantile regression to daily stock returns and market capitalisation data from 100 major European energy companies, while also integrating a set of economic variables and a geopolitical risk index as exogenous drivers. The EFRM index captures major systemic events in the energy markets, namely the 2020 COVID pandemic and the 2022 Russia--Ukraine war. By employing rolling-window VAR and time-varying parameter VAR (TVP-VAR) estimations across monthly and daily frequencies, we document the activation of a previously dormant geopolitical-to-energy-risk transmission channel and the subsequent material attenuation of downstream transmission to European equity and credit markets. Cross-continental validation using US Treasury yields and the US Energy SPDR ETF confirms the global external validity of the EFRM signal. From a practical perspective, the proposed indicator offers policymakers, regulators, and financial stability authorities a high-frequency tool for early detection of systemic threats in one of the most critical components of the EU economy.
- Open Access
- 22 June 2026
- Published by Elsevier in SSRN Electronic Journal
Joint Probabilistic Forecasting for Battery Arbitrage in Day-Ahead Power Markets
Battery energy storage (BESS) is widely considered essential for decarbonising power systems. In this research, we test whether short-term electricity price arbitrage generates sufficient returns to justify investment and BESS economic viability across four interconnected Central and Eastern European day-ahead markets-Romania, Bulgaria, Hungary, and the Czech Republic-from 2018 to 2024, accounting for realistic forecasting uncertainty. All four markets exhibit volatile, heavy-tailed prices and a pronounced daily price seasonality, creating opportunities for arbitrage. We generate joint probabilistic price forecasts using DRF (distributional random forests) and translate these into dayahead battery bidding decisions under real operating constraints. We compare performance with naïve and perfect-foresight benchmarks using standard investment metrics. Empirical findings confirm that probabilistic forecasting adds clear operational value, outperforming a fixed-schedule strategy in every market. The largest gains are seen in the Czech Republic and Bulgaria. Operating a battery in a cross-border setup generates additional profit, driven by persistent regional price differences rather than broad diversification. However, neither forecasting skill nor cross-market flexibility alone is sufficient to make the investment profitable. Therefore, current Li-ion BESS costs make realistic capital recovery unlikely, indicating that deeper market integration, cost reduction and additional revenue streams are necessary for viability.
- 22 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessJoint Probabilistic Forecasting for Battery Arbitrage in Day-Ahead Power Markets
Battery energy storage (BESS) is widely considered essential for decarbonising power systems. In this research, we test whether short-term electricity price arbitrage generates sufficient returns to justify investment and BESS economic viability across four interconnected Central and Eastern European day-ahead markets-Romania, Bulgaria, Hungary, and the Czech Republic-from 2018 to 2024, accounting for realistic forecasting uncertainty. All four markets exhibit volatile, heavy-tailed prices and a pronounced daily price seasonality, creating opportunities for arbitrage. We generate joint probabilistic price forecasts using DRF (distributional random forests) and translate these into dayahead battery bidding decisions under real operating constraints. We compare performance with naïve and perfect-foresight benchmarks using standard investment metrics. Empirical findings confirm that probabilistic forecasting adds clear operational value, outperforming a fixed-schedule strategy in every market. The largest gains are seen in the Czech Republic and Bulgaria. Operating a battery in a cross-border setup generates additional profit, driven by persistent regional price differences rather than broad diversification. However, neither forecasting skill nor cross-market flexibility alone is sufficient to make the investment profitable. Therefore, current Li-ion BESS costs make realistic capital recovery unlikely, indicating that deeper market integration, cost reduction and additional revenue streams are necessary for viability.
- 1 June 2026
- Published by Springer Nature
- p. 95-111
Optimizing Household Energy Use: An Activity-Based Recommendation System for Reducing CO2 Emissions
- 1 June 2026
- Published by Springer Nature
- p. 95-111
Optimizing Household Energy Use: An Activity-Based Recommendation System for Reducing CO2 Emissions
- 7 May 2026
- Published by Springer Nature in Annals of Operations Research
- Vol. 361 (1) , 465-502
Fair models in credit: intersectional discrimination and the amplification of inequity
- 7 May 2026
- Published by Springer Nature in Annals of Operations Research
- Vol. 361 (1) , 465-502
Fair models in credit: intersectional discrimination and the amplification of inequity