All use cases
Use case · Valuation & pricing models

Valuation & pricing

Value, score and price every asset, loan or property on your book in seconds, with models graded on assets they have not seen. Each figure comes with a range and a method you can put in front of a committee, a counterparty or an auditor.

1,544stocks valued · 9 sector models
11 yrpoint-in-time backtest, no look-ahead
22.6%property median error vs asking price, out of sample

When you need this

You need a defensible number for every name on the book, one you can explain when it is challenged. Off-the-shelf templates break on local accounting and market conventions, so coverage ends up thin, inconsistent, and impossible to audit. And a single point estimate hides exactly the information a negotiation or a committee needs: how wrong it could be.

What we build

  • A model for each kind of asset: each sector or asset class gets the method its economics demand.
  • A range on every number: quantile ensembles return p10, p50 and p90, so each figure arrives with its own error bar.
  • Every figure traceable: each one traces to a published formula or a named input, auditable end to end.
  • Methods stress-tested before the market tests them: a frontier reasoning model reviews every revision of the methodology to find where it breaks.
  • Backtests you can believe: point-in-time data only, so no historical figure uses information that was not yet available.
  • Learns from every realised price: each sale or appraisal is scored against its estimate and its range, and the model retrains on those outcomes behind an error gate.
  • Mispricing flagged at portfolio scale: the system shows where price and value diverge.

How it’s built

Your data → sector/asset models (regression · DCF · gradient boosting) → quantile calibration → REST API · screeners · web app

Methods: tabular models & quantile ensembles · classical quantitative models

In our products

vnvalue values 1,544 listed stocks in Vietnam with nine sector models, each figure a published formula on public filings, and re-runs them for every year since 2015 on the data available at the time, publishing the years the model was wrong. vnappraise, a research preview, gives each property estimate a p10 to p90 band and publishes its own error: a 22.6% median error against asking prices (it uses no sale-price data), out of sample over the trailing 7 days, which does not yet meet the 15% bank-collateral bar it published in advance.