System for forecasting expected credit defaults in accordance with IFRS 9 using machine learning and explainable analyses with regulatory audit trails

Abstract

A computer-implemented system (100) for forecasting expected credit losses (ECL) in accordance with IFRS 9 using machine learning and explainable analytics with monitoring audit trails, comprising at least one processor and at least one memory that stores instructions which, when executed, cause the system (100) to operate: a data acquisition interface (1) configured to receive portfolio, customer, risk, repayment, collateral and macroeconomic data from one or more internal and external sources (9); a data management and quality engine (2) configured to perform validation, matching, missing value handling, anomaly detection and data origin marking for the received data; a staging and segmentation engine (3) configured to determine the staging and segmentation in accordance with IFRS 9 for risks based on rules and / or learned risk signals; a feature engineering and feature store module (4) configured to generate, version, and store features that are aligned with observation window and report data; an ECL predictive engine for machine learning (5) configured to train and / or execute predictive models to estimate PD, LGD and EAD and to calculate the 12-month ECL and lifetime ECL under forward-looking scenarios; an explainable analysis engine (6) configured to generate local and global explanation outputs that identify factor contributions to at least staging and ECL outputs, and that produces human-readable explanation codes; an overlay and management customization workflow module (7) configured to apply post-model overlays with documented justifications and approval controls; and a monitoring audit trail and reporting engine (8) configured to generate a timestamped audit record that captures input versions, transformation origin, feature versions, model versions, scenario assumptions, overlays, user actions and report outputs.

Description

INVENTION AREA

[0001] The present invention relates to computer-implemented risk and financial reporting systems, in particular a system (100) for performing forecasts of expected credit losses (ECL) in accordance with IFRS 9 using machine learning, for producing explainable analyses and for maintaining surveillance audit trails for governance, compliance and reproducibility. BACKGROUND OF THE INVENTION

[0002] The subject matter discussed in the "Background" section should not be considered prior art solely because it is mentioned therein. Likewise, a problem mentioned in or related to the subject matter of the "Background" section should not be considered prior art. The subject matter in the "Background" section merely presents various approaches, which could themselves be inventions.

[0003] IFRS 9 requires companies to recognize impairments using an expected credit loss (ECL) approach, which is inherently forward-looking and responds to changes in credit risk over time. Unlike approaches based on actual losses, IFRS 9 requires institutions to estimate expected losses not only based on historical results but also on current conditions and reasonable and verifiable forecasts of future macroeconomic conditions. In practice, this means that institutions must continuously monitor borrower behavior, the dynamics of risk positions, information on collateral, and external macroeconomic influences (e.g.,GDP growth, inflation, unemployment, industry stress indices) must be integrated to achieve ECL results that are consistent, explainable, and defensible in audits and supervisory reviews.

[0004] A major operational challenge arises from the heterogeneity of the data and the challenges related to data quality. Credit risk-relevant data is typically distributed across multiple internal systems, such as core banking systems, credit management, collection platforms, collateral management repositories, and financial / general ledger sources, as well as external sources like credit bureaus and macroeconomic data providers. These datasets often have different schemas, update frequencies, reporting dates, and identifiers. Inconsistent customer / risk keys, missing repayment histories, delayed collateral updates, or conflicting balances between risk and financial systems can significantly distort PD / LGD / EAD estimation and, consequently, ECL.As a result, institutions often have to spend a lot of time and personnel resources reconciling data, manually cleaning up and explaining ECL changes between reporting periods, especially when the underlying cause is a hidden data problem and not a genuine portfolio risk movement.

[0005] Another major challenge is the complexity of classification and SICR measurement. IFRS 9 requires the classification of risk positions into Stage 1, Stage 2, and Stage 3, with ECL measurement based on the 12-month ECL (Stage 1) or the lifetime ECL (Stages 2 and 3), and Stage 3 typically reflecting impaired assets. Determining a significant increase in credit risk (SICR) is rarely based on a single rule but usually involves a combination of default indicators (e.g., days overdue), relative risk changes (e.g., PD migration), watchlist / deferral markers, qualitative triggers, and management guidance.In many institutions, these rules are implemented in isolated spreadsheets, separate risk engines, or fragmented scripts, leading to inconsistent stage results and difficulties in demonstrating to auditors precisely which triggers were activated and why a particular exposure moved between stages. This raises regulatory concerns, especially when the stage migration has a significant impact on provisions.

[0006] Institutions also face technical and methodological challenges in calculating PD, LGD, and EAD across portfolios, particularly when scenario weighting and discounting are required. ECL is not the result of a single model but is typically derived from the interplay of several components: PD maturity structures (12 months and total term), LGD assumptions, which can vary depending on the type of collateral and recovery processes, and EAD, which can depend on utilization behavior, loan conversion factors, repayment schedules, and limits. IFRS 9 further expects the inclusion of several forward-looking macroeconomic scenarios, each weighted with probability, with the final ECL calculated as a scenario-weighted expectation.Furthermore, the discounting of expected payment defaults (often using the effective interest rate) must be applied consistently and reproducibly. These requirements, combined with reporting dates and portfolio segmentation, result in a computationally intensive and governance-sensitive process where even minor changes to assumptions or the timing of data can lead to significant changes in provisions.

[0007] Another obstacle to the practical application of advanced models is the difficulty of explaining the model results, particularly when machine learning techniques are used. Regulators, auditors, and internal model risk committees typically expect institutions to demonstrate that the ECL results are not only statistically reliable but also interpretable and controllable. While machine learning models can improve predictive accuracy, without explanatory capabilities they can be perceived as "black box" risk systems.This leads to recurring audit questions such as: What risk drivers led to an increase in the ECL? Why was a particular account classified as riskier? Does the model behave consistently across all segments? And do macroeconomic inputs appropriately influence the results? If the institution cannot provide justifications at the exposure level and an attribution of drivers at the portfolio level, it will be difficult to defend the ECL analyses during supervisory audits.

[0008] Finally, a regulatory, tamper-proof audit trail that captures complete evidence for the ECL calculation is often lacking. With many conventional implementations, even when model outputs are available, it is difficult to reconstruct the exact conditions of a historical run: Which dataset versions were used, which transformations were applied, which feature definitions were active, which model version and parameters were executed, which scenario assumptions were applied, and whether overlays or manual overrides were introduced. Regulators and auditors increasingly expect the ability to reproduce ECL results for a specific reporting date and to identify every human action in the workflow (author and reviewer approvals, overlay justifications, documentation attachments).Without a robust audit trail and replay package, institutions face increased operational risks, delayed audits, and potential findings related to weaknesses in governance and control.

[0009] Accordingly, there is a clear need for a unified, computerized system that integrates data collection and governance, IFRS 9 staging and SICR determination, machine learning-based PD / LGD / EAD forecasts with scenario-weighted ECL calculation, explainable analytics for supervisory transparency, and tamper-proof audit trails to ensure the full reproducibility and defensibility of ECL results. This need is particularly acute for institutions operating with multiple products and in multiple regions, where scalability, standardization, and supervisory readiness are essential for timely and compliant financial reporting.

[0010] The use of any examples or illustrative phrases (e.g., "such as") in relation to specific embodiments serves only to better illustrate the invention and does not constitute a limitation of the otherwise claimed scope of the invention. No wording in the description shall be construed as referring to an unclaimed element that is essential for carrying out the invention.

[0011] The information disclosed above in this "Background" section is provided solely for a better understanding of the background of the invention and may therefore contain information that is not part of the prior art already known to a person skilled in the art in this country. SUMMARY

[0012] Before describing the systems and methods presented here, it should be noted that this application is not limited to the specific systems and methods described, as there may be several possible embodiments not expressly presented in this disclosure. It should also be noted that the terminology used in the description serves only to describe the specific versions or embodiments and is not intended to limit the scope of this application.

[0013] The present invention relates to a computer-implemented system (100) for forecasting expected credit losses (ECL) in accordance with IFRS 9, which integrates end-to-end data processing, model execution, and monitoring controls in a unified architecture. The system (100) comprises a data ingestion interface (1) for capturing portfolio, customer, risk, repayment, collateral, and macroeconomic data sets from internal and external sources (9), as well as a data management and quality engine (2) for validating, matching, cleansing, and tagging such data sets with origin and version identifiers, thereby ensuring that each reporting run is based on controlled and traceable inputs.

[0014] The system (100) further comprises a staging and segmentation engine (3) for determining IFRS 9 stages and risk segments using configurable rules and / or learned risk signals, as well as a feature engineering and feature store module (4) for generating, standardizing, and versioning features aligned with reporting dates and observation windows for reproducible modeling. An ECL forecasting engine (5) for machine learning trains and / or executes predictive models to estimate PD, LGD, and EAD, and calculates the 12-month and lifetime ECL under forward-looking macroeconomic scenarios, including scenario probability weighting and discounting logic in accordance with the guidelines requirements.

[0015] To support governance, interpretability, and regulatory acceptance, the system (100) includes an explainable analytics engine (6) that generates global and local explanations with justification codes linked to staging and ECL results, and an overlay and management adaptation workflow module (7) that applies documented overlays through maker checker approvals while quantifying their impact. A monitoring, audit trail, and reporting engine (8) captures a timestamped, tamper-proof record of data versions, transformations, staging justifications, feature and model versions, scenario assumptions, overlays, user actions, and final results, and generates auditable evidence packages and replay artifacts to reproduce ECL results for each selected reporting date. BRIEF DESCRIPTION OF THE DRAWING

[0016] To clarify various aspects of some embodiments of the present invention, a more detailed description of the invention is given with reference to specific embodiments illustrated in the accompanying drawing. It is understood that this drawing only illustrates embodiments of the invention and is therefore not to be considered a limitation of its scope. The invention is described and explained with additional specificity and detail using the accompanying drawing.

[0017] To make the advantages of the present invention easily understandable, a detailed description of the invention is given below in conjunction with the accompanying drawing, which, however, should not be regarded as limiting the scope of the invention to the accompanying drawing, in which: Fig. Figure 1 shows a block diagram representation of the system (100) for forecasting expected credit losses in accordance with IFRS 9 using machine learning and explainable analyses with regulatory audit paths. DETAILED DESCRIPTION

[0018] The present invention relates to the system (100) for forecasting expected credit losses in accordance with IFRS 9 using machine learning and explainable analytics with surveillance audit trails.

[0019] Fig. shows a detailed block diagram representation of the system (100) for forecasting expected credit losses in accordance with IFRS 9 using machine learning and explainable analytics with surveillance audit trails.

[0020] With reference to Fig.1 The invention provides a computer-implemented system (100) for forecasting expected credit losses (ECL) in accordance with IFRS 9 using machine learning and explainable analytics with monitoring audit trails. The system (100) is executable on one or more computing devices, including local servers, private / public cloud infrastructure, or hybrid deployments, and comprises at least one processor and at least one memory that stores instructions implementing and coordinating the functional modules (1)-(8). The modules operate in an end-to-end pipeline in which controlled input data is transformed into tiered risks, versioned features, model-based PD / LGD / EAD estimates, scenario-weighted ECL outputs, explainable artifacts, controlled overlays, and auditable supervisory reports, with reproducibility of execution ensured by persistent versioning and event logging.

[0021] The data ingestion interface (1) is configured to capture structured and / or semi-structured data required for IFRS 9 calculations from multiple internal and external sources (9). In practice, the interface (1) receives customer and account master data, risk position balances and transactions, repayment schedules, collection and default indicators (including days overdue), restructuring / deferral flags, collateral attributes and valuation history, information / rating signals, and macroeconomic time series and scenario variables. The interface (1) supports scheduled periodic runs aligned with reporting dates, as well as on-demand capture for ad-hoc analysis, and can implement capture via APIs, secure file transfers, message queues, database connectors, or batch ETL pipelines, assigning run identifiers for downstream traceability.

[0022] The data governance and quality engine (2) standardizes the ingested datasets and enforces quality and governance controls before staging or modeling. The engine (2) performs schema checks, range and threshold validations, reconciliation of balances against authoritative totals such as general ledger or financial / risk control totals, and handling of missing values, duplicates, and anomalous datasets. Furthermore, the engine (2) generates a transformation and origin record that describes the input source, transformation steps, dataset version, reporting date, and quality results, ensuring that any downstream calculation can be linked to a controlled and reproducible dataset state.

[0023] The staging and segmentation engine (3) determines the IFRS 9 stage classification and assigns risks to segments suitable for modeling and reporting. The engine (3) applies configurable policies to assign Stage 1, Stage 2, or Stage 3 and assesses the significant increase in credit risk (SICR) based on one or more indicators such as rules for days overdue, relative risk change (e.g., PD movement), watchlist triggers, qualitative labels, and restructuring / deferral conditions. Furthermore, the engine (3) creates segmentation labels (e.g., product type, region, industry, rating bands, vintage bands, or customer categories) and generates an explicit stage justification for each risk, thereby enabling transparent justification of stage migrations across reporting periods.

[0024] The feature engineering and feature store module (4) generates model-ready features aligned to defined observation windows and the reporting date, and stores these features under strict version control to ensure reproducibility. Module (4) derives behavioral and account dynamics features such as trailing default patterns, utilization trends, repayment volatility, and roll rate characteristics, and can also generate macroeconomically linked features reflecting sensitivity to inflation, interest rates, industry stress, or unemployment movements. Module (4) standardizes and codes features, maintains a governed feature catalog that captures feature definitions, calculation logic, validity dates, and version identifiers, and enables consistent feature extraction during model training, evaluation, validation, and audit repetition.

[0025] The ECL predictive engine for machine learning (5) generates the key IFRS 9 impairment results by estimating PD, LGD, and EAD and calculating 12-month and lifetime ECL values. The engine (5) trains and / or runs one or more predictive models, generates 12-month ECL for Stage 1 risks and lifetime ECL for Stage 2 and Stage 3 risks, and integrates forward-looking information by applying scenario-dependent adjustments driven by macroeconomic variables. The engine (5) calculates the scenario-weighted ECL using probability weights assigned to multiple macroeconomic scenarios and applies discounting (e.g., using the effective interest rate or a configured policy) to expected cash shortfalls while simultaneously generating validation metrics such as backtesting results, calibration summaries, and stability diagnoses for governance.

[0026] The explainable analysis engine (6) generates interpretable results that enable auditors and regulators to understand staging decisions and ECL results, even when machine learning models are used. The engine (6) creates global explanatory views that identify key risk drivers at the portfolio or segment level, as well as local explanatory views that show the contributions of characteristics to the stage, PD, and / or ECL of a given exposure. The engine (6) also generates standardized justification codes and narrative explanations suitable for governance packages and management reviews, and in certain implementations, creates counterfactual explanations that show minimal characteristic changes that would reduce the risk classification or ECL, thereby improving transparency and supporting operational decisions.

[0027] The overlay and management adjustment workflow module (7) enables the controlled application of management overlays and post-model adjustments required for emerging risks, expert opinions, or policy decisions, while ensuring compliance with governance requirements. Module (7) captures the scope of the overlay (portfolio / segment / risk), the rationale, the validity dates, and the accompanying evidence. It enforces approval by the creator and the reviewer using role-based permissions and quantifies the delta impact of the overlays on the ECL at multiple aggregation levels. Module (7) ensures that overlays are applied only after approval and that each overlay remains reversible and traceable, thereby preventing undocumented manual changes and ensuring defensibility to regulatory authorities.

[0028] The regulatory audit chain and reporting engine (8) records complete evidence and generates auditable documentation for each reporting run. The engine (8) logs timestamp events that capture dataset versions and quality results, transformation origins, staging justifications, feature versions, model artifacts and parameters, scenario assumptions and weights, overlay approvals and impacts, user actions, and final results. In one embodiment, the engine (8) stores such records in a tamper-proof, append-only structure (e.g.,chained hashes or equivalent integrity controls) to detect unauthorized changes, generates supervisory report packages including disclosures and voting summaries, and creates an audit replay package that allows the ECL calculation to be rerun for a selected reporting date using the same regulated inputs and versions.

[0029] The figure and the preceding description provide examples of embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements from one embodiment can be added to another embodiment. For example, the sequence of the processes described here can be changed and is not limited to the manner described here. Furthermore, the actions of a block diagram need not be implemented in the sequence shown, nor does it necessarily have to be executed all actions. In addition, those actions that are not dependent on other actions can be executed in parallel with the other actions. The scope of embodiments is by no means limited by these specific examples.

[0030] Although the embodiments of the invention have been described in language relating to structural features and / or methods, it should be noted that the appended claims ( ) are not necessarily limited to the specific features or methods described. Rather, the specific features and methods are disclosed as examples of embodiments of the invention.

Claims

[1] A computer-implemented system (100) for forecasting expected credit losses (ECL) in accordance with IFRS 9 using machine learning and explainable analytics with surveillance audit trails, comprising at least one processor and at least one memory storing instructions which, when executed, cause the system (100) to operate: a data acquisition interface (1) configured to receive portfolio, customer, risk, repayment, collateral and macroeconomic data from one or more internal and external sources (9); a data management and quality engine (2) configured to perform validation, matching, missing value handling, anomaly detection and data origin marking for the received data; a staging and segmentation engine (3) configured to determine the staging and segmentation in accordance with IFRS 9 for risks based on rules and / or learned risk signals; a feature engineering and feature store module (4) configured to generate, version, and store features that are aligned with observation window and report data; an ECL predictive engine for machine learning (5) configured to train and / or execute predictive models to estimate PD, LGD and EAD and to calculate the 12-month ECL and lifetime ECL under forward-looking scenarios; an explainable analysis engine (6) configured to generate local and global explanation outputs that identify factor contributions to at least staging and ECL outputs, and that produces human-readable explanation codes; an overlay and management customization workflow module (7) configured to apply post-model overlays with documented justifications and approval controls; and a monitoring audit trail and reporting engine (8) configured to generate a timestamped audit record that captures input versions, transformation origin, feature versions, model versions, scenario assumptions, overlays, user actions and report outputs. [2] System (100) according to claim 1, wherein the data ingestion interface (1) receives at least one of the following elements: core bank feeds, credit management feeds, bureau and scorecard data, collection data, collateral valuation feeds, IFRS 9 staging indicators and macroeconomic series including GDP, inflation, unemployment and key interest rate. [3] System (100) according to claim 1, wherein the data management and quality engine (2) enforces configurable rules, including threshold checks, referential integrity checks, outlier handling and reconciliation of risk position balances with a general ledger and / or regulatory totals. [4] System (100) according to claim 1, wherein the staging and segmentation engine (3) identifies a significant increase in credit risk (SICR) based on at least one of the following criteria: days overdue, relative PD movement, watchlist / deferral indicators and qualitative trigger flags, and stores a staging justification via the monitoring, audit, trail and reporting engine (8). [5] System (100) according to claim 1, wherein the feature engineering and feature storage module (4) manages feature definitions with version control and enables reproducible feature calculation for a selected report date. [6] System (100) according to claim 1, wherein the ECL forecasting engine (5) for machine learning calculates the scenario-weighted ECL using a plurality of macroeconomic scenarios with associated probability weights and discounts expected payment defaults using an effective interest rate. [7] System (100) according to claim 1, wherein the machine learning ECL forecasting engine (5) performs at least one of the following functions: Cross-validation, backtesting across reporting periods, segment stability tests and calibration of PD outputs to observed failure rates, and logs the results via the monitoring audit trail and reporting engine (8). [8] System (100) according to claim 1, wherein the explainable analysis engine (6) generates at least one of the following functions for selected risk drivers: Shapley-based feature mapping, partial dependency explanations, counterfactual explanations and monotony conformance checks. [9] System (100) according to claim 1, wherein the overlay and management adjustment workflow module (7) implements maker checker controls that require documented justification, attachment of evidence, proof of approval and an automatically calculated delta impact on the ECL at portfolio and segment level. [10] System (100) according to claim 1, wherein the monitoring audit trail and reporting engine (8) stores the audit record in an append-only, tamper-proof structure using chained hashes to detect unauthorized changes and generates an audit rerun package to re-perform the ECL for a selected reporting date.

Images