Use casesFinancial institutions & fintech

  • Asset management company (NPL)
  • Bank
  • Participation bank

Non-performing loan portfolio pricing

A bank's tender gives you as little as ten working days with a data room of enforcement files, mortgage deeds, and appraisals; the bid has to rest on what each file will really collect.

With the systemEvery data-room file read, its recovery estimated from your history, a price range from your rules; the committee bids.

Who decides
Investment committee

Data room · tender · 4 November 2026Illustrative

Title deed record · SME file · encumbrances

First-rank mortgage in favor of another bank, registered 2019. Second-rank mortgage in favor of the selling bank, registered 2021, then an attachment annotation by the enforcement office.

→ second rank behind another bank: priced separately

The document as it arrives. In red, what the system marks for your expert.

1. The work

Today

  1. GatherThe selling bank opens a data room: the loan tape with principal, interest, default date, collateral, and legal status for each debtor, and behind it the enforcement files, mortgage deeds, appraisal reports, and court decisions; analysts sort it by hand.
  2. ReadYour lawyers and analysts read what they can before the deadline: where each enforcement file stands, which rank the mortgage holds, what sits ahead of it, and whether the debtor has died, gone bankrupt, or entered a concordat; the rest is sampled.
  3. EstimateRecovery is judged segment by segment from experience and last year's spreadsheets: how much the unsecured retail files will bring in, and when the mortgaged files will reach a sale.
  4. PriceA cash-flow sheet turns those guesses, your collection and legal costs, and your funding into a price; one wrong assumption on time to sale moves the whole bid.
  5. DecideYour investment committee decides whether to bid, at what price, in cash or as revenue sharing, and which files to leave out.

With the system

  1. GatherThe agent, with read-only access to the data room, splits it into one file per debtor, matches every tape line to its documents, sends each set of documents to a reader, and lists what is missing.Agentic AI
  2. ReadThe SLM reads every enforcement extract, deed, appraisal, and court decision, and returns the legal stage, the collateral and its rank, prior mortgages and seizures, and any death, bankruptcy, or concordat, each tied to its page.SLM
  3. EstimateThe recovery model, trained only on portfolios you have bought and worked, gives each file a band for what it will collect and how long it will take, with reason codes.Model
  4. PriceYour cash-flow and cost rules, run as code, not a model, turn the bands into a price range per segment and show how far it moves if sales through enforcement take longer.Your rule
  5. DecideYour investment committee decides, as before: whether to bid, the price and the structure, and which files to leave out or price apart.Person decides

What lands on the desk

The committee gets every file in the data room already read, each one's recovery band from your own history, and a price range your own cash-flow rules set. The second-rank files are priced apart, and the price is the committee's.

FIG. 1 · Bid fileIllustrative
Tender · data room opened 04.11.2026Unsecured retail and mortgage-backed SME files, one bank

Agent · data room, read-only

  • flaggedTape matched to documentsdata roomsome SME files lack the enforcement extract
  • metPerforming loans in the portfolionone
  • Single risk groupnone
  • Data roomdeleted after the tender, on record

SLM · what the files say

  • Legal stageMost SME files past objection, with seizures on recordenforcement extracts
  • CollateralSeveral SME files hold a second-rank mortgage behind another bankdeed records
  • DebtorOne debtor has died; the heirs are not yet in the fileenforcement file

Model · expected recovery

  • Unsecured retailLow bandslow; most of it after the second year
  • SME, first rankHigher banddepends on the sale through enforcement
  • SME, second rankWide rangepriced apart

Your rules · price range

  • RangeA range per segment under your costs and funding
  • ApartSecond-rank files priced apart
  • SensitivityThe value falls most if sales through enforcement take longer
  • Model
  • SLM
  • Agent
  • Code

Why a person decidesWhether to bid, the price, and the structure are your investment committee's.

DecidesInvestment committee

2. The system

For your technical team — skip it if you like.

How it works

In

  • The seller's loan tape
  • Data-room documents per debtor
  • The tender's terms and calendar

  1. Agent · prepares

    Data-room coordinator

    Orchestrator–workers, read-only

    Reads

    • Everything under In

    Out

    • One file per debtor, documents sorted
    • Each document set sent to a reader
    • Missing documents listed

  2. SLM · understands

    Enforcement and collateral reader

    Fine-tuned on enforcement files, deeds, and appraisals

    Reads

    • Enforcement file extracts
    • Mortgage deeds and land registry records
    • Appraisal reports and court decisions

    Out

    • Legal stage: objection, final, seizure, sale
    • Collateral, its rank, and prior mortgages
    • Death, bankruptcy, or concordat noted
    • Each field tied to its page

  3. Model · estimates

    Recovery model

    Chance of any recovery, then amount; GLM champion, GBM challenger

    Reads

    • The file's fields
    • Your past purchases and what they collected
    • Collateral and legal stage

    Out

    • Recovery band per file
    • Expected time to collect
    • Reason codes
    • Uncertainty per segment

  4. Code · calculates

    Cash-flow and cost rules

    Your pricing sheet and costs, run as code; no model

    Reads

    • Recovery and timing bands
    • Your collection and legal costs
    • Your funding and return hurdle

    Out

    • Price range per segment
    • Sensitivity to time to sale
    • Files to leave out or price apart

  5. Decision

    Investment committee

    Reads

    • The bid file

    Out

    • Bids, or does not bid
    • Sets the price
    • Chooses cash or revenue sharing
    • Leaves files out or prices them apart
How the agent's work is arrangedOrchestrator–workers

Orchestrator–workers

Every data room holds a different mix of documents per debtor, so a coordinator splits it by what each file actually contains, hands each set to a reader, and assembles one bid file.

How the models are trained and tested

Your expert writes the test first. A step the test does not need is skipped, and we say why.

Show the training steps

Small language model (SLM)

  1. Test set

    300–600 files from portfolios you have bought, labeled by your lawyers and analysts: the legal stage, the collateral and its rank, prior mortgages and seizures, and any death, bankruptcy, or concordat.

  2. Dataset

    Only the portfolios you bought, now your own records; a lost tender's data is destroyed under its confidentiality agreement and never trains anything. Debtors' personal data masked in training copies.

  3. Continued pre-training (CPT)If needed

    Only if the base model fails the test on enforcement and land-registry language.

  4. Supervised fine-tuning (SFT)

    On enforcement extracts, mortgage deeds, appraisal reports, and court decisions, scans included.

  5. Reinforcement learning (RL)Skipped

    Skipped: reading fields is taught by fine-tuning and checked against the test set.

Recovery model

  1. Target

    Cash collected per file and when, over the life of portfolios you have bought and worked.

  2. Method and challenger

    Two stages, the chance of any recovery and then the amount; a generalized linear champion against a gradient-boosted challenger, tested on later portfolios.

  3. Validation evidence

    Expected against collected by segment and purchase year, on portfolios already worked out.

  4. Monitoring

    Each quarter, collected against expected per portfolio; the band never prices a file on its own.

Where the agentic system stops

Your investment committee decides whether to bid, at what price, in cash or as revenue sharing, and which files to leave out; at a selling bank, the board's written policy decides what is sold and on what reasoning.

Show its tools, limits, and record
Tools and access
Reads the seller's data room under the tender's confidentiality agreement and your own past purchase and collection records; read-only; writes only the bid file and the list of missing documents.
May not
Set, change, or submit a bid, contact the seller, a debtor, a court, or an enforcement office, copy data-room documents outside the agreed environment, keep a lost tender's data past the agreement, or grant itself any access.
Who decides and the written limit
Investment committeeYour investment committee decides whether to bid, at what price, in cash or as revenue sharing, and which files to leave out; at a selling bank, the board's written policy decides what is sold and on what reasoning. The system never sets or submits a bid, never contacts the seller, a debtor, or a court, and never keeps a lost tender's data past the confidentiality agreement. The recovery model only estimates; it never prices a file on its own.
Stops when
A tape line with no documents, a debtor in more than one risk group, a performing loan in a non-performing portfolio, a collateral it cannot rank, or the deadline near with files unread: the file goes to the analyst, marked.
Record
Every document read, each file's fields and sources, the model and rule versions, the bands and the range, the committee's decision, and the date the data room was deleted.

How agentic systems stay inside limits

3. Value and conditions

Where the time goes, and what your institution will measure.

Where the value comes from

  1. Every file in the data room read before the deadline, not a sample.
  2. Bids built from what your own portfolios collected, file by file.
  3. Lawyer and analyst hours moved to the files that move the price.

What your institution measures

  • Files read before the bid deadline
  • Collected against expected, by portfolio and purchase year
  • Time from data-room opening to committee
  • Files later found with an unread lien or legal issue

Measured in the Decision Blueprint

Your past portfolios' expected and collected cash, and the time from data room to bid, set the baseline; your institution sets the target.

Conditions

The rules, the data, and the known error, with its guard.

  • A bank or another supervised institution sells receivables to an asset-management company only by tender, except a single risk group; the invitation stays on its website at least five working days, every bidder gets the same information, and at least ten working days run from the data sharing to the tender day. The data is shared under a confidentiality agreement, only for the tender, and destroyed after it (BDDK regulation on asset-management companies, Resmî Gazete 14.07.2021 no. 31541, art. 18/2–4, as amended 27.08.2025).
  • The data room holds at least the loan type, the debtor's identity, principal, accrued interest and costs, rate, default date, collateral, any restructuring, the legal proceedings, and whether the loan was taken off the books (art. 18/4). Non-performing loans and loans more than sixty days late may be sold; performing and non-performing loans never in one portfolio, except a single risk group (art. 19/3).
  • The seller writes the economic reason for each sale, including the expected collection and its timing and the fair value, given the debtor and the collateral (art. 18/1); its board sets what is sold under written policies that weigh collection potential, collateral, the debtor's assets, and likely legal outcomes (art. 19/2). The same reader and model serve a selling bank's own file.
  • Real estate an asset-management company takes in a settlement is, as a rule, not sold below the value an independent appraiser set in the year before, nor real estate taken through enforcement below its cost; a sale below it needs its economic reason in writing (art. 21/5); a revenue-sharing contract writes how the seller checks the collections (art. 21/6). These enter your cash-flow rules as written, run as code.

Start here

We take two or three portfolios you have already bought and worked, rebuild what each file collected from your records, collect your pricing sheet and costs, and test the reader on their data-room documents before anything is built.

Who builds it Every engagement is led by an engineer who has built machine-learning models that real decisions run on since 2012, and language models and agentic systems in production since 2024. About eqlx

Request a Decision BlueprintWhat the Decision Blueprint covers

Equalize the gap.

Bring work that takes expertise. Leave knowing the first step.

A short description of the work is enough; no confidential records.

Start with a Decision BlueprintTalk to the engineer

https://eqlx.ai/use-cases/npl-portfolio-pricing · Printed copy; the website is the current version.