vnvalue

The problem

Someone running a portfolio, an advisor or a small fund needs a value it can check for every name it follows: the formula, the inputs and the sector logic behind each figure. Off-the-shelf valuation templates break on local accounting, payout mandates and credit quotas, so coverage ends up thin and inconsistent.

The approach

• A frontier reasoning model is applied recursively to the methodology, stress-testing every revision as a valuation analyst would.
• Nine sector-specific models: private banks on a Huber robust P/B regression on ROE, bank size, a bad-debt proxy and 3-year credit growth (the three large state-owned banks are priced off their peers’ median P/B), utilities on a 3-stage DDM, cyclicals on mid-cycle EV/EBITDA, consumer on an FCF-based DCF, technology on a DCF and P/E blend.
• Per-stock CAPM / WACC: beta from a regression against the VNINDEX, plus a 2.75% country risk premium aligned with Moody’s Ba2 / Fitch BB+. Historical values use point-in-time data only, so the backtest has no look-ahead.
• Forensic screen: a Beneish M-Score ranked by percentile against listed companies in the same year, plus a five-component earnings-quality score weighted by sector. Each week a language model writes a research note on the 100 largest HOSE stocks, from the page’s own model outputs and financial statements.
• A cross-stock owner graph, built with entity resolution, surfaces related-party exposure.

Architecture

Financial data from Vietcap (VCI), with KB Securities as a checked price fallback → 9 sector valuation models → weekly refresh → web app

Results

1,544 stocks valued by nine sector models, refreshed weekly. The track record re-runs the same models for every year since 2015 with only the data available at the time, and it publishes the years the model was wrong. vnvalue calls its output consistent, reproducible and fully disclosed, and makes no claim that the model is correct.