Credit-risk and provision models built for audit scrutiny
AlgoSolution builds expected-credit-loss (ECL), provision, borrowing-base, and liquidity models — and the data platforms that feed them — for regulated lenders in Canada and the United States.
Every credit modeled as its own probability of loss — millions of rows, so the provision reflects the full loss distribution, not one cohort average.
Per-credit information beats cohort averages
Standard IFRS 9 approaches average whole cohorts. Modeling each credit's probability of loss uses the granular information the cohort average throws away — and yields provisions grounded in per-credit evidence, not averages.
| Dimension | Cohort averages | Loan-level probabilistic |
|---|---|---|
| Information used | Aggregate cohort behavior | Every credit's own characteristics and history |
| Provision precision | Coarse; mix shifts distort estimates | Estimates built from per-credit information — robust to portfolio-mix shifts that distort cohort averages |
| Transparency | One number, limited insight into drivers | Exposes the reasonable range of ECL outcomes and its sensitivities |
| Scale | Spreadsheet-bound; breaks on large tapes | Engineered for millions of loan-level rows |
The models — and the platform they stand on
Loan-Level Risk Modeling
Cohort-average methods discard the per-credit information sitting in a loan tape. AlgoSolution models each credit's probability of loss from its own characteristics and history — on a warehouse, ETL, and SQL-modeling platform engineered for millions of loan-level rows, the layer where most modeling efforts quietly fail — yielding provisions grounded in per-credit evidence and robust to the portfolio-mix shifts that distort cohort averages.
Deliverables: a loan-level risk model, the data platform that feeds it, and documented model logic and assumptions.
IFRS 9 & CECL Provisions
Allowance frameworks demand a number management can defend to auditors and regulators. AlgoSolution builds transparent provision models for IFRS 9 and CECL (ASC 326) that expose the reasonable range of ECL outcomes and its sensitivities, so management can choose a methodology consistent with its accounting policy and risk tolerance — within a defensible range.
Deliverables: IFRS 9 / CECL provision calculations, the reasonable range and its sensitivities, and a defensible methodology write-up.
Validation & Performance Monitoring
A provision is only credible if its accuracy is measured, not assumed. AlgoSolution confronts forecast losses with realized losses on an ongoing basis and delivers monitoring dashboards for default rates, charge-off rates, and financial-statement variances — so tracking error is a measured property that can be shown to independent reviewers.
Deliverables: realized-versus-forecast tracking, monitoring dashboards, and a periodic performance review.
The same loan tape also drives borrowing-base eligibility and liquidity forecasts →
Tested where it counts: under audit, in production
Selected engagement — a Canadian fintech consumer lender
Built and refined the loan-loss provision model that drives the lender's balance-sheet and income-statement provisions and expected-loss estimates — a model that has withstood challenge from multiple external audit firms across six annual audit cycles. The balance-sheet provision is estimated as lifetime charge-offs net of recoveries on the existing book, projected from current-book behavior. Alongside it: a borrowing-base / collateral financing-structure model, the data warehouse and ETL pipeline that feed both, and the reporting dashboards management uses to monitor default rates, charge-off rates, and the financial statements.
Probabilistic, transparent, and tracked against reality
Probability distributions, not point guesses
Expected loss is estimated from full probability distributions at the credit level — a discipline drawn from probabilistic modeling, applied to lending portfolios. The same loan-level machinery measures lifetime expected losses under IFRS 9 and CECL (ASC 326) alike.
Transparency by construction
The model shows its drivers, assumptions, and the sensitivity of the ECL range to each — management sees why the number is what it is, and what would move it.
Realized-vs-forecast tracking
Forecasts are confronted with realized losses on an ongoing basis. New signals fold in over time — per-loan repayment-cycle tracking and recovery patterns on charged-off loans — so tracking error is measured and, cycle over cycle, decreasing.
Frequently asked questions
How is this different from our current IFRS 9 cohort approach?
Cohort approaches average behavior across buckets of loans, which discards the per-credit information sitting in your loan tape. Loan-level probabilistic modeling uses that information directly — each credit's own characteristics and history — which yields expected-loss estimates built from per-credit evidence rather than cohort averages, and makes the drivers of the number visible.
Will the model stand up to our external auditors?
The reference engagement's provision model has withstood challenge from multiple external audit firms across six annual audit cycles, while driving a lender's actual balance-sheet and income-statement provisions. Monthly net charge-offs have tracked to within ±$20K on roughly $2.5M realized-vs-forecast. Tracking error is measured monthly against realized charge-offs net of recoveries; the figure is the current tracking level, not a guarantee. Transparency is built in: the reasonable range of ECL outcomes and its sensitivities are exposed, which is precisely what independent reviewers probe.
What data do we need to have in place?
A loan-level tape — per-credit records of characteristics, balances, and payment history. It does not need to be clean or centralized: building the warehouse, ETL, and SQL modeling layer that makes the tape usable is part of the service, and engagements routinely start there. Production work has run at roughly 15 million loan-level rows.
Can management choose the provisioning methodology?
Yes — that is the point of transparency. The model exposes the reasonable range of ECL outcomes and the sensitivity of that range to assumptions, and management chooses a methodology consistent with its accounting policy and risk tolerance, within a defensible range it can explain to auditors and the board.
How does an engagement start, and what does it cost?
It starts with a scoping conversation about your portfolio, framework, and constraints. There is no published price — scope drives the engagement, and you will know exactly what is proposed before anything begins.
Methodology notes
ECL methodology: exposing the reasonable range
Why a transparent range beats a black-box point estimate — and how management chooses within it.
Read the noteLiquidity & cash-flow forecasting from loan-level history
Turning historical payment patterns into liquidity and financial-statement forecasts.
Read the noteWorking with loan-level data at scale
What breaks at millions of rows, and the engineering that keeps per-credit modeling feasible.
Read the noteRelated capabilities
This specialist application combines several core capabilities of AlgoSolution’s practice around one set of lender requirements: financial modeling for the provision and statement logic, data engineering for the loan-level platform, business analytics for the monitoring dashboards, and data science & AI for the probabilistic methods underneath.
Financial Modeling · Data Engineering · Business Analytics · Data Science & AI · Financing & Borrowing Base