{"paper_id":"2410.02846","title":"A Spatio-Temporal Machine Learning Model for Mortgage Credit Risk: Default Probabilities and Loan Portfolios","authors":["Pascal Kündig","Fabio Sigrist"],"paper_date":"2024-10-03","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2410.02846v3","page_url":"https://thequant.space/flowcharts/a-spatio-temporal-machine-learning-model-for-mortgage-credit/","scores":{"empirical_rigor":8.5,"math_complexity":9,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a novel machine learning model for credit risk that combines tree-boosting with a latent spatio‑temporal Gaussian process accounting for frailty correlation.","quote":"We introduce a novel machine learning model for credit risk by combining tree-boosting with a latent spatio-temporal Gaussian process model accounting for frailty correlation.","summary":"We introduce a novel machine learning model for credit risk by combining tree-boosting with a latent spatio-temporal Gaussian process model accounting for frailty correlation. This allows for modeling non-linearities and interactions among predictor variables in a flexible data-driven manner and…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"A loan is considered in default if it is at least 90 days delinquent.","type":"data","stated":"explicit","quote":"a loan is considered to be in default if it is at least 90 days delinquent."},{"assumption":"The integral in the likelihood cannot be calculated in closed form and is approximated using a Laplace approximation.","type":"model","stated":"explicit","quote":"First, the integral in (6) cannot be calculated in closed form, and we approximate it using a Laplace approximation."},{"assumption":"Vecchia approximations are used to obtain a sparse reverse Cholesky factor for the latent Gaussian process.","type":"model","stated":"explicit","quote":"calculate an approximate sparse reverse Cholesky factor of the precision matrix of the latent Gaussian process b."},{"assumption":"Training data consists of all loan‑year observations up to and including 2007, with predictions made for 2008, then expanding the window year by year.","type":"data","stated":"explicit","quote":"we first consider all loan-year observations up to and including the year 2007 as training data and make predictions for all loan-year observations of 2008."},{"assumption":"Tuning parameters are selected to maximize AUC on a validation set consisting of the most recent year in the training window.","type":"statistical","stated":"explicit","quote":"... selecting the combinations of tuning parameters that maximize the area under the receiver operating characteristic curve (AUC) on the validation data."}]},"datasets":{"status":"available","items":[{"name":"Freddie Mac single‑family loan‑level data set (release 37)","provider":"Freddie Mac","kind":"alternative","access":"public","period":"1999‑2022","universe":"union of all fully amortizing 30‑year fixed‑rate mortgages in random subsets","quote":"We consider the union of all fully amortizing 30‑year fixed‑rate mortgages in all random subsets of Freddie Mac."}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"weekly-monthly"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R3","candidates":{"front_matter":"n/a","evidence_card":"n/a","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"forecast-accuracy","findings":[{"finding":"Predictive default probabilities and loan portfolio loss distributions from the tree‑boosted spatio‑temporal model are more accurate than those from conventional independent linear hazard models and linear spatio‑temporal models.","kind":"predictive-accuracy","quote":"we find that both predictive default probabilities for individual loans and predictive loan portfolio loss distributions obtained with our novel approach are more accurate compared to conventional independent linear hazard models and also linear spatio‑temporal models."},{"finding":"The tree‑boosted spatial model achieves the lowest CRPS (2.034 × 10⁸) and RMSE (3.823 × 10⁸) among all models, indicating best performance for the center of portfolio loss distributions.","kind":"performance","quote":"the tree‑boosted spatial model has the lowest CRPS and RMSE, i.e., this model is the most accurate for the center of portfolio loss distributions."},{"finding":"The tree‑boosted spatio‑temporal model has the lowest 99 % quantile loss (1.021 × 10⁸), providing the most accurate predictive upper quantiles of portfolio loss distributions.","kind":"performance","quote":"the tree‑boosted spatio‑temporal model has the lowest quantile loss. In other words, this model yields the most accurate predictive upper quantiles of portfolio loss distributions."},{"finding":"The RMSE difference between the tree‑boosted spatial model and the linear spatial model is approximately US $35 million.","kind":"performance","quote":"the difference in RMSE between the tree‑boosted spatial model and the linear spatial model are approximately US $35 million."},{"finding":"During the global financial crisis (2008), the tree‑boosted spatial and spatio‑temporal frailty models predict realistic mean portfolio losses, whereas linear models overestimate losses.","kind":"predictive-accuracy","quote":"the tree‑boosted spatial and spatio‑temporal frailty models clearly predict the most realistic mean portfolio losses during the global financial crisis around the year 2008. The linear models clearly overestimate the portfolio loss."}],"reported_performance":null,"out_of_sample":{"method":"walk-forward","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":17,"citations":{"cited_by_in_corpus":2,"by_independent_authors":2,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":2,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":true,"code_url":"https://github.com/pkuendig/SpaceTimeML","note":null},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":1,"extracted_at":"2026-11-14T18:00:46Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}