EstimateMachine-learning models
Who gets credit and at what limit, which customer is about to leave, which machine will stop, what price to offer. A model produces the probability and its reasons; thresholds your policy writes tie each band to an action; the action runs in your own systems, and its result is measured.
FIG. 1 · Decision chain · SME limit reviewIllustrativeStrategy in force
One account, from score to the action applied.
Band 6Behavioral score · band 3 a year ago
Every account outside a written cell goes to the credit manager.
When it fits, and when you don't need us
A chance of default in credit scoring, a failure risk, a forecast, a rank: something nobody reads off the file alone, and a rule no longer separates.
Repaid or defaulted, failed or ran: how past cases turned out is what the model learns from.
You don't need us
If a written rule or a lookup already gives the right answer, use code.
Not sure which kind fits your decision? The Decision Blueprint answers that first.
How we build it
We do not stop at the score. We design the thresholds and the action matrix with you, simulate them on your history, connect them to your own systems, and measure the result.
1 · Model
Default, leaving, paying, buying, failing. Learned from how your past cases turned out, tested on cases held back.
2 · Thresholds
Your institution writes the limit. We simulate the options on your history and show approval and loss side by side.
3 · Action matrix
Approve, limit, call, offer, payment plan, review. The matrix is written, and every cell has its rule and its cap.
4 · Applied
Your decision engine, CRM, or limit service applies it. Every case outside a written cell goes to a named person.
5 · Measured
Each action's result is compared with a control group, and we tune the thresholds on the evidence, together.
The result is the evidence for the next threshold change; your institution approves every change.
Example · the sales side
The model ranks who is likely to leave; the matrix says what to do by the customer's value. A customer who will stay gets no needless discount.
Your institution writes the discount cap, the channels, and the contact frequency; the model only ranks.
FIG. 2 · Action matrix · churn risk × customer valueIllustrative
| Value: high | Value: medium | Value: low | |
|---|---|---|---|
| Risk: high | A relationship manager calls; a retention offer inside the approval limit | A tailored offer, through a digital channel | A service reminder; no discount |
| Risk: medium | Onto the relationship manager's calendar | Into the campaign segment | No action |
| Risk: low | No action: a customer who will stay gets no discount | ||
What we build
Find your model by the result you need. Each row shows what the model produces and the action a threshold ties it to.
For:BanksConsumer finance companiesFactoringLeasingInsurancePayment and e-money institutionsManufacturingServices
Who gets credit, how much, on what terms; who is deteriorating, whom to call.
| Model | What it produces | Action, by threshold | Where |
|---|---|---|---|
| Application scorecardUse case | Chance of default at application, with decline reasons | Automatic approval · grey-zone review · decline | Banks, consumer finance, leasing |
| Behavioral score and limit managementUse case | A monthly behavioral band and its migration | Limit increase · freeze the undrawn limit · drawdown approval | Banks, consumer finance, cards |
| Early-warning modelUse case | Chance of deterioration six to twelve months ahead | Watch list · interim financials requested · committee file | Banks, factoring, leasing |
| Collections scoreUse case | Chance of self-cure and of rolling further into arrears | Reminder · call · payment plan · pre-legal review | Banks, consumer finance, factoring |
| SME and corporate ratingUse case | Internal grade, chance of default, and the reasons it moved | Authority level · collateral condition · review frequency | Banks, factoring, leasing |
| IFRS 9 expected credit loss | Probability of default (PD), loss given default (LGD), and exposure at default (EAD) | Input to your provisioning; the staging decision stays yours | Banks, finance companies, leasing, factoring |
| Risk-based pricingUse case | A cost floor and the chance of acceptance at each price | Recommended price · relationship manager's discretion band | Banks, consumer finance, leasing |
| Fraud scoreUse case | Chance of fraud on an application or a transaction | Close on a written pattern · review · stop | Banks, payment institutions |
| Factoring: buyer and invoice risk | Chance the buyer pays, and a duplicate-invoice signal | Buyer limit · discount rate · review | Factoring |
| Insurance: claims fraud and severityUse case | A fraud band and the expected cost | Fast track · loss adjuster · special investigation | Insurance |
Who is leaving, what sells to whom, which price is accepted.
| Model | What it produces | Action, by threshold | Where |
|---|---|---|---|
| Churn model | Chance the customer leaves, and why | A retention action by the customer's value | Banks, insurance, payments, services |
| Customer segmentation | A few meaningful groups, from behavior | Campaign design and channel by group | Every sector |
| Propensity to buy (sales model)Use case | Chance of buying, product by product | Whom to offer, through which channel, and when | Banks, insurance, finance companies |
| Next best offerUse case | One offer per customer, with its reasons | Relationship manager's queue · contact-frequency cap | Banks, insurance |
| Price elasticity and price offerUse case | Chance of acceptance and expected profit at each price | The most profitable price inside your policy's limits | Finance, insurance renewals, manufacturers' quotes |
| Campaign effect (uplift) | The customers an action is most likely to change | Contact the customers most likely to move | Banks, insurance, services |
When it will stop, how much will be needed, where the scrap comes from.
| Model | What it produces | Action, by threshold | Where |
|---|---|---|---|
| Predictive maintenanceUse case | Risk of an unplanned stop in the coming days, and the signals that moved | Draft work order · parts reserved | Manufacturing |
| Quality and yieldUse case | Chance of scrap per batch, and its drivers | A proposed process window · a test plan | Manufacturing |
| Demand forecasting | Demand by product and period | Stock and production plan | Manufacturing, distribution |
| Workload and capacity forecasting | Volume of calls, applications, and files | Shift and team plan | Services, finance operations |
Your institution writes the threshold and the action. No method is presumed better; the method is chosen on your data.
Detail · the price offer
The same logic works for a loan rate, an insurance renewal, and a manufacturer's quote. Protected attributes never enter the price, and we test the inputs for proxies of them.
FIG. 3 · Price response · one segmentIllustrative
If you already have a model
Most institutions already run models. We re-evaluate a model against what it is for, then take the step the evidence calls for: sometimes recalibration is enough, sometimes it needs redevelopment.
When there is no model, or a new product or segment opens.
model/ · pipeline/
When ranking power has dropped or the portfolio has changed.
evaluation-report
When the ranking holds but the probabilities no longer match outcomes.
calibration report
Tying the score to cut-offs and an action matrix.
strategy matrix
A new model or strategy beside the current one, on a small share.
test results
Drift, calibration, and the results of the actions, with a recalibration route.
monitoring/
Intended use, data, method, limits, and test results, in the form your unit reads.
documentation/
The model and the strategy wired into your own engine, on your own servers.
runbooks/
We prepare the evidence; your independent validation unit judges it.
No method is presumed better, and one test metric never picks the production method on its own.
Calibration is one of the gates: when the model says one in ten, about one in ten such cases should turn out that way.
FIG. 4 · CalibrationIllustrative
Calibration: predicted against observed
One gate of several. Discrimination, stability, and reason codes are tested separately before independent validation.
What you receive
FIG. 5 · HandoverIllustrative
model/The model and its codepipeline/The data pipeline, with its sourcesevaluation-reportRanking power, calibration, and stability on held-back casesdocumentation/Intended use, limits, and methodstrategy/The thresholds and the action matrix, with their simulation on your historymonitoring/Monitoring of drift and calibration, with a recalibration routerunbooks/How your team runs it, with trainingOn your own servers or in your cloud account. Regulated work stays on your premises, cut off from outside networks where required.
Your data trains your model only, and never another customer's system.
Desks that use it
Application scorecard for new-to-bank applicants
TodayMost applicants are new to you, the bureau file is thin, and your only outcomes are for the people you approved.
With the systemA credit scoring model beside today's rules: a scorecard, with the default probability by band.
Credit policy committee decides
Behavioral score and limit management
TodayLimits are set once at application; good customers outgrow them and quietly deteriorating ones keep full, undrawn lines.
With the systemLimit actions queued with their evidence, inside the committee's matrix.
Underwriter decides
Early warning on the loan book
TodayThe borrower is current; other banks are already cutting its limits, and you see its quarter only at renewal.
With the systemA file, not a spreadsheet: signals checked in parallel and gathered in one place.
Watch-list committee decides
Risk-based pricing with price response
TodayYour price covers risk and capital, yet you lose good customers over a few basis points and discount those who would have signed anyway.
With the systemA price grid that shows how each segment responds to a different rate.
Pricing committee decides
TodayMost alerts are false, and each one still takes an analyst real time to rule out.
With the systemEach alert arrives with one note; sanctions hits stay with a person.
Fraud analyst decides
TodayScrap is found at the end of the line, and nobody knows which setting caused it.
With the systemThe setting most likely behind the scrap, with a window to test.
Process engineer decides
It works on its own, and stacks with:
Equalize the gap.
Tell us the decision, the outcomes on record, and what is done with the score today. The Decision Blueprint shows whether a model, a strategy, or both is the right first step.
https://eqlx.ai/ml-models · Printed copy; the website is the current version.