No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Multimodal Machine Learning for Credit Modeling
Abstract
Credit ratings are traditionally generated using models that use financial statement data and market data, which is tabular (numeric and categorical). Practitioner and academic models do not include text data. Using an automated approach to combine long-form text from SEC filings with the tabular data, we show how multimodal machine learning using stack ensembling and bagging can generate more accurate rating predictions. This paper demonstrates a methodology to use big data to extend tabular data models, which have been used by the ratings industry for decades, to the class of multimodal machine learning models.
... In this work, we will develop a novel end-to-end machine learning system, called Fake Advertisements Detection using Automated Multimodal Learning (FADAML), to detect and filter out fake online advertisements. Our system combines techniques in multimodal machine learning [2,3] and automated machine learning (AutoML) [4,5] to improve the system's detection rate. To our knowledge, this is the first work that combines multimodal and automated machine learning for fake advertisement detection. ...
... Such multimodal machine learning systems [2] have been gaining popularity in recent years due to the availability of different data types such as texts, tabular data, images, sounds, etc., for each application. Previous works also combined multimodal machine learning and AutoML [3,47] for text and tabular features. In [47], several AutoML systems were compared on 18 different multimodal datasets that contain both texts and hand-crafted features. ...
... In [47], several AutoML systems were compared on 18 different multimodal datasets that contain both texts and hand-crafted features. In [3], AutoML was used for credit modeling on financial data that combines raw texts from Securities and Exchange Commission filings with features extracted by financial experts. In these works, multimodal AutoML can achieve high accuracies while minimizing the implementation efforts. ...
The popularity of e-commerce has given rise to fake advertisements that can expose users to financial and data risks while damaging the reputation of these e-commerce platforms. For these reasons, detecting and removing such fake advertisements are important for the success of e-commerce websites. In this paper, we propose FADAML, a novel end-to-end machine learning system to detect and filter out fake online advertisements. Our system combines techniques in multimodal machine learning and automated machine learning to achieve a high detection rate. As a case study, we apply FADAML to detect fake advertisements on popular Vietnamese real estate websites. Our experiments show that we can achieve 91.5% detection accuracy, which significantly outperforms three different state-of-the-art fake news detection systems.
... Golbayani et al. (2020a), Huang et al. (2004), Lee (2007) and Wallis et al. (2019) compared various statistical and machine learning methods for prediction. Existing research in this domain also achieves higher accuracy for credit rating prediction by incorporating various ML techniques like feature selection (Hajek & Michalak, 2013), balancing minority classes (Hájek & Olej, 2014), ensembling schemes (Abellán & Castellano, 2017;Hájek & Olej, 2014;Marqués et al., 2012;Nguyen et al., 2021), and fusion models (Wu & Hsu, 2012). Further, Golbayani et al. (2020a) provides an extensive comparative study of corporate credit rating prediction models. ...
... To the best of our knowledge, corporate credit rating datasets for US companies are either paid or proprietary (Golbayani et al., 2020a;Hajek & Michalak, 2013;Hájek & Olej, 2014;Kim, 2005;Martens et al., 2009). Open datasets are collectively large (Nguyen et al., 2021), but filtering those datasets for a particular business sector and a particular year makes the dataset too small. This poses a major challenge for performing detailed temporal analysis for a single business sector. ...
... • Training data window size Huang et al. (2004), Lee (2007), Nguyen et al. (2021), Novotná (2012) and Wu and Hsu (2012) combined the credit rating data of all available years and randomly divided it into a training-testing dataset. This traditional method for obtaining a training-testing dataset is not coherent with the credit rating prediction problem because of an absolute requirement to test the model for samples that are outside the time window of the training data. ...
Specialized agencies issue corporate credit ratings to evaluate the creditworthiness of a company, serving as a crucial financial indicator for potential investors. These ratings offer a tangible understanding of the risks associated with the credit investment returns of a company. Every company aims to achieve a favorable credit rating, as it enables them to attract more investments and reduce their cost of capital. Credit rating agencies typically employ unique rating scales that are broadly categorized into investment-grade or non-investment-grade (junk) classes. Given the extensive assessment conducted by credit rating agencies, it becomes a challenge for companies to formulate a straightforward and all-encompassing set of rules which may help to understand and improve their credit rating. This paper employs explainable AI, specifically decision trees, using historical data to establish an empirical rule on financial ratios. The rule obtained using the proposed approach can be effectively utilized to understand as well as plan and attain an investment-grade rating. Additionally, the study investigates the temporal aspect by identifying the optimal time window for training data. As the availability of structured data for temporal analysis is currently limited, this study addresses this challenge by creating a large and high-quality curated dataset. This dataset serves as a valuable resource for conducting comprehensive temporal analysis. Our analysis demonstrates that the empirical rule derived from historical data, yields a high precision value, and therefore highlights the effectiveness of our proposed approach as a valuable guideline and a feasible decision support system.
... In this work, we will develop a novel end-to-end machine learning system, called Fake Advertisements Detection using Automated Multimodal Learning (FADAML), to detect and filter out fake online advertisements. Our system combines techniques in multimodal machine learning [2,3] and automated machine learning (AutoML) [4,5] to improve the system's detection rate. To our knowledge, this is the first work that combines multimodal and automated machine learning for fake advertisement detection. ...
... Such multimodal machine learning systems [2] have been gaining popularity in recent years due to the availability of different data types such as texts, tabular data, images, sounds, etc., for each application. Previous works also combined multimodal machine learning and AutoML [3,47] for text and tabular features. In [47], several AutoML systems were compared on 18 different multimodal datasets that contain both texts and hand-crafted features. ...
... In [47], several AutoML systems were compared on 18 different multimodal datasets that contain both texts and hand-crafted features. In [3], AutoML was used for credit modeling on financial data that combines raw texts from Securities and Exchange Commission filings with features extracted by financial experts. In these works, multimodal AutoML can achieve high accuracies while minimizing the implementation efforts. ...
The popularity of e-commerce has given rise to fake advertisements that can expose users to financial and data risks while damaging the reputation of these e-commerce platforms. For these reasons, detecting and removing such fake advertisements are important for the success of e-commerce websites. In this paper, we propose FADAML, a novel end-to-end machine learning system to detect and filter out fake online advertisements. Our system combines techniques in multimodal machine learning and automated machine learning to achieve a high detection rate. As a case study, we apply FADAML to detect fake advertisements on popular Vietnamese real estate websites. Our experiments show that we can achieve 91.5% detection accuracy, which significantly outperforms three different state-of-the-art fake news detection systems.
... While a number of studies have begun to explore this textual modality, it's fair to say that it has not been fully capitalized on in CCR analysis. These studies typically rely on traditional feature extraction techniques like sentiment analysis, N-gram models, bag-of-words, and document embeddings (e.g., doc2vec), which may not capture semantic and syntactic information as effectively as modern deep learning approaches (Choi et al., 2020;Nguyen et al., 2021;Tsai & Wang, 2017;Wang et al., 2023). In terms of multimodal fusion, the prevalent approaches involve simply concatenating features from different modalities without considering the inherent interconnections between them (Mai et al., 2019;Stevenson et al., 2021). ...
... As this study focuses on CCR prediction, which poses a binary classification problem, the data are bifurcated into good (above BB) and junk labels in CCR-S, following the approach of Nguyen et al. (2021). In contrast, CCR-L adopts a BB+ threshold for labeling as per Makwana et al. (2022). ...
Precise corporate credit risk (CCR) prediction empowers investors, banks and other financial institutions to build risk prevention and crisis evasion mechanisms. Currently, CCR prediction gradually trends towards integrating multi-source data for more accurate prediction, but the effective fusion of these data is still insufficient. For this purpose, this paper presents a novel method titled the multimodal extended deep belief network (MEDBN) for CCR prediction. First, a multimodal dataset comprising numerical, categorical, and textual data is constructed based on two public corporate credit rating datasets, with textual data sourced from 10-K/Q filings. Each modality is processed using specialized preprocessing techniques. Then, to effectively extract features from each modality while addressing the challenges of multimodal fusion and joint representation learning, the DBN is extended by incorporating advanced deep learning techniques, including residual networks (ResNet), embedding layers, and bidirectional encoders (BERT). Finally, MEDBN undergoes two-stage training consisting of unsupervised pretraining followed by supervised fine-tuning, enabling sufficient learning of joint representations to enhance model performance. Extensive experimental results indicate that MEDBN consistently outperforms benchmark models across most key metrics, with particularly notable robustness on smaller datasets. These findings imply that MEDBN can effectively fuse and utilize multimodal data to achieve more reliable predictions. Additionally, this study demonstrates how the textual content influences performance, providing interpretable insights to support the decision-making process.
... The surge in machine learning and its applications has opened up a new arena of possibilities in diverse fields ranging from image recognition, natural language processing to finance [1], [2], [3], [4], [5]. However, a major challenge for building or training predictive models is the scarcity of labelled data [6], [7]. ...
... [4] create a dictionary of Environmental and Social (E&S) phrases, while [42] leveraged word-embeddings to measure the corporate culture. Moreover, multimodal machine learning was used by [2] and [43] for credit rating prediction and measurement of persuasiveness respectively. [1] investigated biases in the multimodal analysis of financial earnings calls. ...
p>In the past few years, Transformer based models have shown excellent performance across a variety of tasks and domains. However, the black-box nature of these models, along with their high computing and manual annotation costs have limited adoption of these models. In this paper, we employ a weak-supervision-based approach to alleviate these concerns. We build and compare models for financial claim detection task using sentences with numerical information in analyst reports for more than 1500 public companies in the United States from 2017 to 2020. The proposed aggregation function for the weak-supervision model outperforms existing aggregation functions. In addition to standard performance metrics, we provide cost-value analysis of human-annotation and weak-supervision labeling along with estimates of the carbon footprint of our models. We also analyze the performance of our claim detection models across various industry sectors given the considerable variation in numerical financial claims across industries. Our work highlights the potential of weak supervision models for research at the intersection of Finance and Data Engineering. </p
... The surge in machine learning and its applications has opened up a new arena of possibilities in diverse fields ranging from image recognition, natural language processing to finance [1], [2], [3], [4], [5]. However, a major challenge for building or training predictive models is the scarcity of labelled data [6], [7]. ...
... [4] create a dictionary of Environmental and Social (E&S) phrases, while [42] leveraged word-embeddings to measure the corporate culture. Moreover, multimodal machine learning was used by [2] and [43] for credit rating prediction and measurement of persuasiveness respectively. [1] investigated biases in the multimodal analysis of financial earnings calls. ...
p>In the past few years, Transformer based models have shown excellent performance across a variety of tasks and domains. However, the black-box nature of these models, along with their high computing and manual annotation costs have limited adoption of these models. In this paper, we employ a weak-supervision-based approach to alleviate these concerns. We build and compare models for financial claim detection task using sentences with numerical information in analyst reports for more than 1500 public companies in the United States from 2017 to 2020. The proposed aggregation function for the weak-supervision model outperforms existing aggregation functions. In addition to standard performance metrics, we provide cost-value analysis of human-annotation and weak-supervision labeling along with estimates of the carbon footprint of our models. We also analyze the performance of our claim detection models across various industry sectors given the considerable variation in numerical financial claims across industries. Our work highlights the potential of weak supervision models for research at the intersection of Finance and Data Engineering. </p
... Recent literature has shown that adding text based measures can improve a variety of prediction tasks in the finance domain. Nguyen et al. (2021) show that adding text-based features along with accounting variables improve predictions of a firm's credit rating. Tetlock, Saar-Tsechansky and Macskassy (2008) analyze how quantitative measures of text are useful in predicting individual firms' performance. ...
Using a comprehensive sample of 2,585 bankruptcies from 1990 to 2019, we benchmark the performance of various machine learning models in predicting financial distress of publicly traded U.S. firms. We find that gradient boosted trees outperform other models in one-year-ahead forecasts. Variable permutation tests show that excess stock returns, idiosyncratic risk, and relative size are the more important variables for predictions. Textual features derived from corporate filings do not improve performance materially. In a credit competition model that accounts for the asymmetric cost of default misclassification, the survival random forest is able to capture large dollar profits.
We propose two frameworks to deal with problem settings in which both structured and unstructured data are available. Structured data problems are best solved by traditional machine learning models such as boosting and tree-based algorithms, whereas deep learning has been widely applied to problems dealing with images, text, audio, and other unstructured data sources. However, for the setting in which both structured and unstructured data are accessible, it is not obvious what the best modeling approach is to enhance performance on both data sources simultaneously. Our proposed frameworks allow joint learning on both kinds of data by integrating the paradigms of boosting models and deep neural networks. The first framework, the boosted-feature-vector deep learning network, learns features from the structured data using gradient boosting and combines them with embeddings from unstructured data via a two-branch deep neural network. Secondly, the two-weak-learner boosting framework extends the boosting paradigm to the setting with two input data sources. We present and compare first- and second-order methods of this framework. Our experimental results on both public and real-world datasets show performance gains achieved by the frameworks over selected baselines by magnitudes of 0.1%–4.7%.
fastai is a deep learning library which provides practitioners with high-level components that can quickly and easily provide state-of-the-art results in standard deep learning domains, and provides researchers with low-level components that can be mixed and matched to build new approaches. It aims to do both things without substantial compromises in ease of use, flexibility, or performance. This is possible thanks to a carefully layered architecture, which expresses common underlying patterns of many deep learning and data processing techniques in terms of decoupled abstractions. These abstractions can be expressed concisely and clearly by leveraging the dynamism of the underlying Python language and the flexibility of the PyTorch library. fastai includes: a new type dispatch system for Python along with a semantic type hierarchy for tensors; a GPU-optimized computer vision library which can be extended in pure Python; an optimizer which refactors out the common functionality of modern optimizers into two basic pieces, allowing optimization algorithms to be implemented in 4–5 lines of code; a novel 2-way callback system that can access any part of the data, model, or optimizer and change it at any point during training; a new data block API; and much more. We used this library to successfully create a complete deep learning course, which we were able to write more quickly than using previous approaches, and the code was more clear. The library is already in wide use in research, industry, and teaching.
This monograph surveys the technology and empirics of text analytics in finance. I present various tools of information extraction and basic text analytics. I survey a range of techniques of classification and predictive analytics, and metrics used to assess the performance of text analytics algorithms. I then review the literature on text mining and predictive analytics in finance, and its connection to networks, covering a wide range of text sources such as blogs, news, web posts, corporate filings, etc. I end with textual content presenting forecasts and predictions about future directions.
MXNet is a multi-language machine learning (ML) library to ease the
development of ML algorithms, especially for deep neural networks. Embedded in
the host language, it blends declarative symbolic expression with imperative
tensor computation. It offers auto differentiation to derive gradients. MXNet
is computation and memory efficient and runs on various heterogeneous systems,
ranging from mobile devices to distributed GPU clusters.
This paper describes both the API design and the system implementation of
MXNet, and explains how embedding of both symbolic expression and tensor
operation is handled in a unified fashion. Our preliminary experiments reveal
promising results on large scale deep neural network applications using
multiple GPU machines.
The inherent nature of social media content poses serious challenges to practical applications of sentiment analysis. We present VADER, a simple rule-based model for general sentiment analysis, and compare its effectiveness to eleven typical state-of-practice benchmarks including LIWC, ANEW, the General Inquirer, SentiWordNet, and machine learning oriented techniques relying on Naive Bayes, Maximum Entropy, and Support Vector Machine (SVM) algorithms. Using a combination of qualitative and quantitative methods, we first construct and empirically validate a gold-standard list of lexical features (along with their associated sentiment intensity measures) which are specifically attuned to sentiment in microblog-like contexts. We then combine these lexical features with consideration for five general rules that embody grammatical and syntactical conventions for expressing and emphasizing sentiment intensity. Interestingly , using our parsimonious rule-based model to assess the sentiment of tweets, we find that VADER outperforms individual human raters (F1 Classification Accuracy = 0.96 and 0.84, respectively), and generalizes more favorably across contexts than any of our benchmarks.
We present a method for constructing ensembles from libraries of thousands of models. Model libraries are generated using different learning algorithms and parameter settings. Forward stepwise selection is used to add to the ensemble the models that maximize its performance. Ensemble selection allows ensembles to be optimized to performance metric such as accuracy, cross entropy, mean precision, or ROC Area. Experiments with seven test problems and ten metrics demonstrate the benefit of ensemble selection.
This paper proposes a new tree-based ensemble method for supervised classification and regression problems. It essentially consists of randomizing strongly both attribute and cut-point choice while splitting a tree node. In the extreme case, it builds totally randomized trees whose structures are independent of the output values of the learning sample. The strength of the randomization can be tuned to problem specifics by the appropriate choice of a parameter. We evaluate the robustness of the default choice of this parameter, and we also provide insight on how to adjust it in particular situations. Besides accuracy, the main strength of the resulting algorithm is computational efficiency. A bias/variance analysis of the Extra-Trees algorithm is also provided as well as a geometrical and a kernel characterization of the models induced.
This article presents convenient reduced-form models of the valuation of contingent claims subject to default risk, focusing
on applications to the term structure of interest rates for corporate or sovereign bonds. Examples include the valuation of
a credit-spread option.
We present a method for constructing ensembles from libraries of thousands of models.
The inherent nature of social media content poses serious challenges to practical applications of sentiment analysis. We present VADER, a simple rule-based model for general sentiment analysis, and compare its effectiveness to eleven typical state-of-practice benchmarks including LIWC, ANEW, the General Inquirer, SentiWordNet, and machine learning oriented techniques relying on Naive Bayes, Maximum Entropy, and Support Vector Machine (SVM) algorithms. Using a combination of qualitative and quantitative methods, we first construct and empirically validate a gold-standard list of lexical features (along with their associated sentiment intensity measures) which are specifically attuned to sentiment in microblog-like contexts. We then combine these lexical features with consideration for five general rules that embody grammatical and syntactical conventions for expressing and emphasizing sentiment intensity. Interestingly, using our parsimonious rule-based model to assess the sentiment of tweets, we find that VADER outperforms individual human raters (F1 Classification Accuracy = 0.96 and 0.84, respectively), and generalizes more favorably across contexts than any of our benchmarks.
Textual analysis, implemented at scale, has become an important addition to the methodological toolbox of finance. In this review, given the proliferation of papers now using this method, we first provide an updated survey of the literature while focusing on a few broad topics—social media, political bias, and detecting fraud. We do not attempt to survey the various statistical methods and instead initially focus on the construction and use of lexicons in finance. We then center the discussion on readability as an attribute frequently incorporated in contemporaneous research, arguing that its use begs the question of what we are measuring. Finally, we discuss how the literature might build on the intent of measuring readability to measure something more appropriate and more broadly relevant—complexity.
An ever-increasing share of human interaction, communication, and culture is recorded as digital text. We provide an introduction to the use of text as an input to economic research. We discuss the features that make text different from other forms of data, offer a practical overview of relevant statistical methods, and survey a variety of applications. (JEL C38, C55, L82, Z13)
This paper investigates the impact of a firm’s annual report readability and ambiguous tone on its borrowing costs. We find that firms with larger 10-K file sizes and a higher proportion of uncertain and weak modal words in 10-Ks have stricter loan contract terms and greater future stock price crash risk. Our results suggest that the readability and tone ambiguity of a firm’s financial disclosures are related to managerial information hoarding. Shareholders of firms with less readable and more ambiguous annual reports not only suffer from less transparent information disclosure but also bear the increased cost of external financing.
We propose a new measure of readability, the Bog Index, which captures the plain English attributes of disclosure (e.g., active voice, fewer hidden verbs, etc.). We validate this measure using a series of controlled experiments and an archival-based regulatory intervention to prospectus filing readability. We also demonstrate the importance of understanding the underlying drivers of quantity-based measures of readability. In particular, we caution researchers that a vast amount of the variation in Form 10-K file size over time is driven by the inclusion of content unrelated to the underlying text in the 10-K (e.g., HTML, XML, PDFs).
Tree boosting is a highly effective and widely used machine learning method. In this paper, we describe a scalable end-to-end tree boosting system called XGBoost, which is used widely by data scientists to achieve state-of-the-art results on many machine learning challenges. We propose a novel sparsity-aware algorithm for sparse data and weighted quantile sketch for approximate tree learning. More importantly, we provide insights on cache access patterns, data compression and sharding to build a scalable tree boosting system. By combining these insights, XGBoost scales beyond billions of examples using far fewer resources than existing systems.
Tree boosting is a highly effective and widely used machine learning method. In this paper, we describe a scalable end-to-end tree boosting system called XGBoost, which is used widely by data scientists to achieve state-of-the-art results on many machine learning challenges. We propose a novel sparsity-aware algorithm for sparse data and weighted quantile sketch for approximate tree learning. More importantly, we provide insights on cache access patterns, data compression and sharding to build a scalable tree boosting system. By combining these insights, XGBoost scales beyond billions of examples using far fewer resources than existing systems.
This article provides a new methodology for pricing and hedging derivative securities involving credit risk. Two types of credit risks are considered. The first is where the asset underlying the derivative security may default. The second is where the writer of the derivative security may default. We apply the foreign currency analogy of Jarrow and Turnbull (1991) to decompose the dollar payoff from a risky security into a certain payoff and a "spot exchange rate." Arbitrage-free valuation techniques are then employed. This methodology can be applied to corporate debt and over the counter derivatives, such as swaps and caps. © 2008 by World Scientific Publishing Co. Pte. Ltd. All rights reserved.
Measuring the extent to which a firm is financially constrained is critical in assessing capital structure. Extant measures of financial constraints focus on macro firm characteristics such as age and size, variables highly correlated with other firm attributes. We parse 10-K disclosures filed with the U.S. Securities and Exchange Commission (SEC) using a unique lexicon based on constraining words. We find that the frequency of constraining words exhibits very low correlation with traditional measures of financial constraints and predicts subsequent liquidity events, such as dividend omissions or increases, equity recycling, and underfunded pensions, better than widely used financial constraint indexes.
This article provides a new methodology for pricing and hedging derivative securities involving credit risk. Two types of credit risks are considered. The first is where the asset underlying the derivative security may default. The second is where the writer of the derivative security may default. We apply the foreign currency analogy of Jarrow and Turnbull (1991) to decompose the dollar payoff from a risky security into a certain payoff and a ''spot exchange rate.'' Arbitrage-free valuation techniques are then employed. This methodology can be applied to corporate debt and over the counter derivatives, such as swaps and caps.
Random forests are a combination of tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. The generalization error for forests converges a.s. to a limit as the number of trees in the forest becomes large. The generalization error of a forest of tree classifiers depends on the strength of the individual trees in the forest and the correlation between them. Using a random selection of features to split each node yields error rates that compare favorably to Adaboost (Y. Freund & R. Schapire, Machine Learning: Proceedings of the Thirteenth International conference, ***, 148–156), but are more robust with respect to noise. Internal estimates monitor error, strength, and correlation and these are used to show the response to increasing the number of features used in the splitting. Internal estimates are also used to measure variable importance. These ideas are also applicable to regression.
Among the simplest and most intuitively appealing classes of nonprobabilistic classification procedures are those that weight the evidence of nearby sample observations most heavily. More specifically, one might wish to weight the evidence of a neighbor close to an unclassified observation more heavily than the evidence of another neighbor which is at a greater distance from the unclassified observation. One such classification rule is described which makes use of a neighbor weighting function for the purpose of assigning a class to an unclassified sample. The admissibility of such a rule is also considered.
Models of financial distress rely primarily on accounting-based information (e.g. [Altman, E., 1968. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. Journal of Finance 23, 589–609; Ohlson, J., 1980. Financial ratios and the probabilistic prediction of bankruptcy. Journal of Accounting Research 19, 109–131]) or market-based information (e.g. [Merton, R.C., 1974. On the pricing of corporate debt: The risk structure of interest rates. Journal of Finance 29, 449–470]). In this paper, we provide evidence on the relative performance of these two classes of models. Using a sample of 2860 quarterly CDS spreads we find that a model of distress using accounting metrics performs comparably to market-based structural models of default. Moreover, a model using both sources of information performs better than either of the two models. Overall, our results suggest that both sources of information (accounting- and market-based) are complementary in pricing distress.
If options are correctly priced in the market, it should not be possible to make sure profits by creating portfolios of long and short positions in options and their underlying stocks. Using this principle, a theoretical valuation formula for options is derived. Since almost all corporate liabilities can be viewed as combinations of options, the formula and the analysis that led to it are also applicable to corporate liabilities such as common stock, corporate bonds, and warrants. In particular, the formula can be used to derive the discount that should be applied to a corporate bond because of the possibility of default.
In this paper we conduct a systematic comparison of confidence intervals around estimated probabilities of default (PD) using several analytical approaches as well as parametric and nonparametric bootstrap methods. We do so for two different PD estimation methods, cohort and duration (intensity), with 22 years of credit ratings data. We find that the bootstrapped intervals for the duration-based estimates are relatively tight when compared to either analytic or bootstrapped intervals around the less efficient cohort estimator. We show how the large differences between the point estimates and confidence intervals of these two estimators are consistent with non-Markovian migration behavior. Surprisingly, even with these relatively tight confidence intervals, it is impossible to distinguish notch-level PDs for investment grade ratings, e.g. a PDAA− from a PDA+. However, once the speculative grade barrier is crossed, we are able to distinguish quite cleanly notch-level estimated PDs. Conditioning on the state of the business cycle helps: it is easier to distinguish adjacent PDs in recessions than in expansions.
AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data
- N Erickson
- J Mueller
- A Shirkov
- H Zhang
- P Larroy
- M Li
CatBoost: Unbiased boosting with categorical features
- L Prokhorenkova
- G Gusev
- A Vorobev
- A V Dorogush
- A Gulin
LightGBM: A highly efficient gradient boosting decision tree
- G Ke
- Q Meng
- T Finley
- T Wang
- W Chen
- W Ma
AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data
- erickson
CatBoost: Unbiased boosting with categorical features
- prokhorenkova
LightGBM: A highly efficient gradient boosting decision tree
- ke