Pub Date : 2025-01-16DOI: 10.1016/j.ijforecast.2024.07.004
Doron Sonsino , Yefim Roth
A large panel of chief financial officers’ forecasts of the S&P 500 annual returns and four experiments suggest that forecast confidence decreases as the forecasts diverge from zero, in the positive or negative direction. This decreased confidence is reflected in longer forecast intervals, larger perceived volatility estimates, and weaker belief in the accuracy of the predictions. De Bondt’s (1993) forecast hedging intensifies with the extremity of the forecasts, but the decrease in confidence is sustained when the intervals are symmetrized. Imposing cumulative prospect theory preferences on the CFOs, permutation tests show that the decreased confidence delays the response to optimistic expectations and alleviates miscalibration, although the optimistic CFOs still discount the VIX by more than 50%. The paper thus reveals a self-corrective mechanism that partially, but far from fully, offsets the overconfidence hazards.
{"title":"The decrease in confidence with forecast extremity","authors":"Doron Sonsino , Yefim Roth","doi":"10.1016/j.ijforecast.2024.07.004","DOIUrl":"10.1016/j.ijforecast.2024.07.004","url":null,"abstract":"<div><div>A large panel of chief financial officers’ forecasts of the S&P 500 annual returns and four experiments suggest that forecast confidence decreases as the forecasts diverge from zero, in the positive or negative direction. This decreased confidence is reflected in longer forecast intervals, larger perceived volatility estimates, and weaker belief in the accuracy of the predictions. <span><span>De Bondt</span></span>’s (<span><span>1993</span></span>) forecast hedging intensifies with the extremity of the forecasts, but the decrease in confidence is sustained when the intervals are symmetrized. Imposing cumulative prospect theory preferences on the CFOs, permutation tests show that the decreased confidence delays the response to optimistic expectations and alleviates miscalibration, although the optimistic CFOs still discount the VIX by more than 50%. The paper thus reveals a self-corrective mechanism that partially, but far from fully, offsets the overconfidence hazards.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 3","pages":"Pages 877-893"},"PeriodicalIF":6.9,"publicationDate":"2025-01-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144212027","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-26DOI: 10.1016/j.ijforecast.2024.07.006
Robert Fildes , Paul Goodwin , Shari De Baets
Forecast value added (FVA) analysis is commonly used to measure the improved accuracy and bias achieved by judgmentally modifying system forecasts. Assessing the factors that prompt such adjustments, and their effect on forecast performance, is important in demand forecasting and planning. To address these issues, we collected the publicly available data on around 147,000 forecasts from six studies and analysed them using a common framework. Adjustments typically led to improvements in bias and accuracy for only just over half of stock keeping units (SKUs), though there was variation across datasets. Positive adjustments were confirmed as more likely to worsen performance. Negative adjustments typically led to improvements, particularly when they were large. The evidence that forecasters made effective use of relevant information not available to the algorithm was weak. Instead, they appeared to respond to irrelevant cues, or those of less diagnostic value. The key question is how organizations can improve on their current forecasting processes to achieve greater forecast value added. For example, a debiasing procedure applied to adjusted forecasts proved effective at improving forecast performance.
{"title":"Forecast value added in demand planning","authors":"Robert Fildes , Paul Goodwin , Shari De Baets","doi":"10.1016/j.ijforecast.2024.07.006","DOIUrl":"10.1016/j.ijforecast.2024.07.006","url":null,"abstract":"<div><div>Forecast value added (FVA) analysis is commonly used to measure the improved accuracy and bias achieved by judgmentally modifying system forecasts. Assessing the factors that prompt such adjustments, and their effect on forecast performance, is important in demand forecasting and planning. To address these issues, we collected the publicly available data on around 147,000 forecasts from six studies and analysed them using a common framework. Adjustments typically led to improvements in bias and accuracy for only just over half of stock keeping units (SKUs), though there was variation across datasets. Positive adjustments were confirmed as more likely to worsen performance. Negative adjustments typically led to improvements, particularly when they were large. The evidence that forecasters made effective use of relevant information not available to the algorithm was weak. Instead, they appeared to respond to irrelevant cues, or those of less diagnostic value. The key question is how organizations can improve on their current forecasting processes to achieve greater forecast value added. For example, a debiasing procedure applied to adjusted forecasts proved effective at improving forecast performance.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 2","pages":"Pages 649-669"},"PeriodicalIF":6.9,"publicationDate":"2024-12-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143579226","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-25DOI: 10.1016/j.ijforecast.2024.11.011
Audronė Virbickaitė , Hedibert F. Lopes , Martina Danielova Zaharieva
This article investigates the benefits of combining information available from daily and intraday data in financial return forecasting. The two data sources are combined via a density pooling approach, wherein the individual densities are represented as a copula function, and the potentially time-varying pooling weights depend on the forecasting performance of each model. The dependence structure in the daily frequency case is extracted from a standard static and dynamic conditional covariance modeling, and the high-frequency counterpart is based on a realized covariance measure. We find that incorporating both high- and low-frequency information via density pooling provides significant gains in predictive model performance over any individual model and any model combination within the same data frequency. A portfolio allocation exercise quantifies the economic gains by producing investment portfolios with the smallest variance and highest Sharpe ratio.
{"title":"Multivariate dynamic mixed-frequency density pooling for financial forecasting","authors":"Audronė Virbickaitė , Hedibert F. Lopes , Martina Danielova Zaharieva","doi":"10.1016/j.ijforecast.2024.11.011","DOIUrl":"10.1016/j.ijforecast.2024.11.011","url":null,"abstract":"<div><div>This article investigates the benefits of combining information available from daily and intraday data in financial return forecasting. The two data sources are combined via a density pooling approach, wherein the individual densities are represented as a copula function, and the potentially time-varying pooling weights depend on the forecasting performance of each model. The dependence structure in the daily frequency case is extracted from a standard static and dynamic conditional covariance modeling, and the high-frequency counterpart is based on a realized covariance measure. We find that incorporating both high- and low-frequency information via density pooling provides significant gains in predictive model performance over any individual model and any model combination within the same data frequency. A portfolio allocation exercise quantifies the economic gains by producing investment portfolios with the smallest variance and highest Sharpe ratio.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 3","pages":"Pages 1184-1198"},"PeriodicalIF":6.9,"publicationDate":"2024-12-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144211939","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-20DOI: 10.1016/j.ijforecast.2024.11.010
Luc Bauwens , Yongdeng Xu
Realized variance–covariance models define the conditional expectation of a realized variance–covariance matrix as a function of past matrices using a GARCH-type structure. We evaluate the forecasting performance of various models in terms of economic value, measured through economic loss functions, across two datasets. Our empirical findings reveal that models incorporating realized volatilities offer significant economic value and outperform GARCH models relying solely on daily returns for daily and weekly horizons. Among two realized variance–covariance measures, using a directly rescaled intraday measure for full-day estimation provides more economic value than employing overnight returns, which tends to produce noisier estimators of overnight covariance, diminishing their predictive effectiveness. The HEAVY-H model for the variance–covariance matrix of the daily return demonstrates superior or comparable performance to the best-performing realized variance–covariance models, making it a preferred choice for empirical analysis.
{"title":"The contribution of realized variance–covariance models to the economic value of volatility timing","authors":"Luc Bauwens , Yongdeng Xu","doi":"10.1016/j.ijforecast.2024.11.010","DOIUrl":"10.1016/j.ijforecast.2024.11.010","url":null,"abstract":"<div><div>Realized variance–covariance models define the conditional expectation of a realized variance–covariance matrix as a function of past matrices using a GARCH-type structure. We evaluate the forecasting performance of various models in terms of economic value, measured through economic loss functions, across two datasets. Our empirical findings reveal that models incorporating realized volatilities offer significant economic value and outperform GARCH models relying solely on daily returns for daily and weekly horizons. Among two realized variance–covariance measures, using a directly rescaled intraday measure for full-day estimation provides more economic value than employing overnight returns, which tends to produce noisier estimators of overnight covariance, diminishing their predictive effectiveness. The HEAVY-H model for the variance–covariance matrix of the daily return demonstrates superior or comparable performance to the best-performing realized variance–covariance models, making it a preferred choice for empirical analysis.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 3","pages":"Pages 1165-1183"},"PeriodicalIF":6.9,"publicationDate":"2024-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144211938","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-20DOI: 10.1016/j.ijforecast.2024.11.009
Andrej Sokol
I propose a new model, conditional quantile regression (CQR), that generates density forecasts consistent with a specific view of the future evolution of some of the explanatory variables. This addresses a shortcoming of existing quantile regression-based models in settings that require forecasts to be conditional on technical assumptions, such as most forecasting processes within policy institutions. Through an application to house price inflation in the euro area, I show that CQR provides a viable alternative to conditional density forecasting with Bayesian VARs, with added flexibility and further insights that do not come at the cost of forecasting performance.
{"title":"Fan charts 2.0: Flexible forecast distributions with expert judgement","authors":"Andrej Sokol","doi":"10.1016/j.ijforecast.2024.11.009","DOIUrl":"10.1016/j.ijforecast.2024.11.009","url":null,"abstract":"<div><div>I propose a new model, conditional quantile regression (CQR), that generates density forecasts consistent with a specific view of the future evolution of some of the explanatory variables. This addresses a shortcoming of existing quantile regression-based models in settings that require forecasts to be conditional on technical assumptions, such as most forecasting processes within policy institutions. Through an application to house price inflation in the euro area, I show that CQR provides a viable alternative to conditional density forecasting with Bayesian VARs, with added flexibility and further insights that do not come at the cost of forecasting performance.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 3","pages":"Pages 1148-1164"},"PeriodicalIF":6.9,"publicationDate":"2024-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144211937","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-10DOI: 10.1016/j.ijforecast.2024.11.005
David A. Comerford , Jack B. Soll
A majority of Americans reported the economy to be worsening when objective indicators showed it to be recovering. We show that this is symptomatic of attribute substitution—people answer a taxing question as though asked a related easy-to-answer question. An implication of attribute substitution is that forecasts will vary across a direct format, which asks whether the economy will be better in 12 months, versus an indirect format, which asks respondents to rate both current conditions and the conditions they expect for 12 months’ time. We compare these formats in three studies and over 2,000 respondents. Relative to the direct format, the indirect format delivers trends that show greater consensus across Republicans and Democrats; are less equivocal about the course of the US economy; and are more realistic about the magnitude of change in opinion poll data.
{"title":"Partisan bias, attribute substitution, and the benefits of an indirect format for eliciting forecasts and judgments of trend","authors":"David A. Comerford , Jack B. Soll","doi":"10.1016/j.ijforecast.2024.11.005","DOIUrl":"10.1016/j.ijforecast.2024.11.005","url":null,"abstract":"<div><div>A majority of Americans reported the economy to be worsening when objective indicators showed it to be recovering. We show that this is symptomatic of attribute substitution—people answer a taxing question as though asked a related easy-to-answer question. An implication of attribute substitution is that forecasts will vary across a direct format, which asks whether the economy will be better in 12 months, versus an indirect format, which asks respondents to rate both current conditions and the conditions they expect for 12 months’ time. We compare these formats in three studies and over 2,000 respondents. Relative to the direct format, the indirect format delivers trends that show greater consensus across Republicans and Democrats; are less equivocal about the course of the US economy; and are more realistic about the magnitude of change in opinion poll data.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 2","pages":"Pages 702-715"},"PeriodicalIF":6.9,"publicationDate":"2024-12-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143579228","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-10DOI: 10.1016/j.ijforecast.2024.08.003
Giacomo Sbrana , Andrea Silvestrini
The Theta method is a well-established prediction benchmark widely used in forecast competitions. This method has received significant attention since it was introduced more than 20 years ago, with several authors proposing variants to improve its performance. This paper considers multiple sources of error versions for Theta, belonging to the family of structural time series models. It investigates its out-of-sample forecast performance using the extensive M4-Competition dataset, which includes 100,000 time series. We compare the proposed structural Theta model against several benchmarks, including all variants of the Theta method. The results demonstrate its remarkable predictive abilities as it outperforms all its variants and competitors, emerging as a solid benchmark for use in forecast competitions.
{"title":"The structural Theta method and its predictive performance in the M4-Competition","authors":"Giacomo Sbrana , Andrea Silvestrini","doi":"10.1016/j.ijforecast.2024.08.003","DOIUrl":"10.1016/j.ijforecast.2024.08.003","url":null,"abstract":"<div><div>The Theta method is a well-established prediction benchmark widely used in forecast competitions. This method has received significant attention since it was introduced more than 20 years ago, with several authors proposing variants to improve its performance. This paper considers multiple sources of error versions for Theta, belonging to the family of structural time series models. It investigates its out-of-sample forecast performance using the extensive M4-Competition dataset, which includes 100,000 time series. We compare the proposed structural Theta model against several benchmarks, including all variants of the Theta method. The results demonstrate its remarkable predictive abilities as it outperforms all its variants and competitors, emerging as a solid benchmark for use in forecast competitions.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 3","pages":"Pages 940-952"},"PeriodicalIF":6.9,"publicationDate":"2024-12-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144211931","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-06DOI: 10.1016/j.ijforecast.2024.09.004
Joscha Beckmann , Robert L. Czudaj
This paper provides a new perspective on the expectations-building mechanism in foreign exchange markets. We analyze the role of expectations regarding macroeconomic fundamentals for expected exchange rate changes. Real-time survey data is assessed for 29 economies from 2002 to 2023, and expectations regarding GDP growth, inflation, interest rates, and current accounts are considered. Our empirical findings show that fundamentals expectations are more important over longer than shorter horizons. We find that an expected increase in GDP growth relative to the US leads to an expected appreciation of the domestic currency. In contrast, higher relative inflation expectations lead to an expected depreciation, a finding consistent with purchasing power parity. Our results also indicate that the expectation-building process differs systematically across pessimistic and optimistic forecasts, with the former paying more attention to fundamentals expectations. Finally, we also observe that fundamentals expectations have some explanatory power for forecast errors, especially for longer horizons.
{"title":"Fundamental determinants of exchange rate expectations","authors":"Joscha Beckmann , Robert L. Czudaj","doi":"10.1016/j.ijforecast.2024.09.004","DOIUrl":"10.1016/j.ijforecast.2024.09.004","url":null,"abstract":"<div><div>This paper provides a new perspective on the expectations-building mechanism in foreign exchange markets. We analyze the role of expectations regarding macroeconomic fundamentals for expected exchange rate changes. Real-time survey data is assessed for 29 economies from 2002 to 2023, and expectations regarding GDP growth, inflation, interest rates, and current accounts are considered. Our empirical findings show that fundamentals expectations are more important over longer than shorter horizons. We find that an expected increase in GDP growth relative to the US leads to an expected appreciation of the domestic currency. In contrast, higher relative inflation expectations lead to an expected depreciation, a finding consistent with purchasing power parity. Our results also indicate that the expectation-building process differs systematically across pessimistic and optimistic forecasts, with the former paying more attention to fundamentals expectations. Finally, we also observe that fundamentals expectations have some explanatory power for forecast errors, especially for longer horizons.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 3","pages":"Pages 1003-1021"},"PeriodicalIF":6.9,"publicationDate":"2024-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144211934","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-06DOI: 10.1016/j.ijforecast.2024.11.004
Alessio Brini , Giacomo Toscano
This paper introduces SpotV2Net, a multivariate intraday spot volatility forecasting model based on a graph attention network architecture. SpotV2Net represents assets as nodes within a graph and includes non-parametric high-frequency Fourier estimates of the spot volatility and co-volatility as node features. Further, it incorporates Fourier estimates of the spot volatility of volatility and co-volatility of volatility as features for node edges, to capture spillover effects. We test the forecasting accuracy of SpotV2Net in an extensive empirical exercise, conducted with the components of the Dow Jones Industrial Average index. The results we obtain suggest that SpotV2Net yields statistically significant gains in forecasting accuracy, for both single-step and multi-step forecasts, compared to a panel heterogeneous autoregressive model and alternative machine-learning models. To interpret the forecasts produced by SpotV2Net, we employ GNNExplainer (Ying et al., 2019), a model-agnostic interpretability tool, and thereby uncover subgraphs that are critical to a node’s predictions.
介绍了基于图关注网络结构的多变量日内现货波动率预测模型SpotV2Net。SpotV2Net将资产表示为图中的节点，并将现货波动率和协同波动率的非参数高频傅立叶估计作为节点特征。此外，它将波动率的现货波动率和波动率的协同波动率的傅立叶估计作为节点边缘的特征，以捕获溢出效应。我们用道琼斯工业平均指数的组成部分进行了广泛的实证练习，测试了SpotV2Net的预测准确性。我们获得的结果表明，与面板异构自回归模型和替代机器学习模型相比，SpotV2Net在单步和多步预测方面的预测准确性在统计上有显著提高。为了解释SpotV2Net产生的预测，我们使用了gnexplainer (Ying et al., 2019)，这是一种与模型无关的可解释性工具，从而揭示了对节点预测至关重要的子图。
{"title":"SpotV2Net: Multivariate intraday spot volatility forecasting via vol-of-vol-informed graph attention networks","authors":"Alessio Brini , Giacomo Toscano","doi":"10.1016/j.ijforecast.2024.11.004","DOIUrl":"10.1016/j.ijforecast.2024.11.004","url":null,"abstract":"<div><div>This paper introduces SpotV2Net, a multivariate intraday spot volatility forecasting model based on a graph attention network architecture. SpotV2Net represents assets as nodes within a graph and includes non-parametric high-frequency Fourier estimates of the spot volatility and co-volatility as node features. Further, it incorporates Fourier estimates of the spot volatility of volatility and co-volatility of volatility as features for node edges, to capture spillover effects. We test the forecasting accuracy of SpotV2Net in an extensive empirical exercise, conducted with the components of the Dow Jones Industrial Average index. The results we obtain suggest that SpotV2Net yields statistically significant gains in forecasting accuracy, for both single-step and multi-step forecasts, compared to a panel heterogeneous autoregressive model and alternative machine-learning models. To interpret the forecasts produced by SpotV2Net, we employ GNNExplainer (Ying et al., 2019), a model-agnostic interpretability tool, and thereby uncover subgraphs that are critical to a node’s predictions.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 3","pages":"Pages 1093-1111"},"PeriodicalIF":6.9,"publicationDate":"2024-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144212024","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-05DOI: 10.1016/j.ijforecast.2024.07.005
Yufei Xia , Zhiyin Han , Yawen Li , Lingyun He
Fintech lending experiences high credit risk and needs an efficient credit scoring model, but it also faces limited data sources and severe class imbalance. We develop a novel two-stage credit scoring model (called LLM-FP-CatBoost) by solving the two issues simultaneously. Large language models (LLMs) initially extract narrative data as a supplementary credit dataset. A new FocalPoly loss is then incorporated with CatBoost to handle the class imbalance problem. Extensive comparisons demonstrate that the proposed LLM-FP-CatBoost significantly outperforms the benchmarks in most circumstances. When making pairwise comparisons between LLMs on the fintech lending dataset, we found that the Chinese-specific LLM, i.e., ERNIE 4.0, achieves the best overall performance, followed by GPT-4 and BERT-based models. The performance decomposition reveals that the superiority is mainly attributed to the new data source extracted by the LLMs. The SHAP algorithm further ensures the interpretability of LLM-FP-CatBoost. The superiority of the proposed LLM-FP-CatBoost model remains robust to hyperparameters of the loss function, specific LLMs, and other extraction methods of narrative data. Finally, we discuss some managerial implications concerning credit scoring in fintech lending.
{"title":"Credit scoring model for fintech lending: An integration of large language models and FocalPoly loss","authors":"Yufei Xia , Zhiyin Han , Yawen Li , Lingyun He","doi":"10.1016/j.ijforecast.2024.07.005","DOIUrl":"10.1016/j.ijforecast.2024.07.005","url":null,"abstract":"<div><div>Fintech lending experiences high credit risk and needs an efficient credit scoring model, but it also faces limited data sources and severe class imbalance. We develop a novel two-stage credit scoring model (called LLM-FP-CatBoost) by solving the two issues simultaneously. Large language models (LLMs) initially extract narrative data as a supplementary credit dataset. A new FocalPoly loss is then incorporated with CatBoost to handle the class imbalance problem. Extensive comparisons demonstrate that the proposed LLM-FP-CatBoost significantly outperforms the benchmarks in most circumstances. When making pairwise comparisons between LLMs on the fintech lending dataset, we found that the Chinese-specific LLM, i.e., ERNIE 4.0, achieves the best overall performance, followed by GPT-4 and BERT-based models. The performance decomposition reveals that the superiority is mainly attributed to the new data source extracted by the LLMs. The SHAP algorithm further ensures the interpretability of LLM-FP-CatBoost. The superiority of the proposed LLM-FP-CatBoost model remains robust to hyperparameters of the loss function, specific LLMs, and other extraction methods of narrative data. Finally, we discuss some managerial implications concerning credit scoring in fintech lending.</div></div>","PeriodicalId":14061,"journal":{"name":"International Journal of Forecasting","volume":"41 3","pages":"Pages 894-919"},"PeriodicalIF":6.9,"publicationDate":"2024-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144212028","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"经济学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号